Get live statistics and analysis of Stefano Ermon's profile on X / Twitter
AI Prof @Stanford | CEO & Cofounder @_inception_ai
| Co-inventor of DDIM, FlashAttention, DPO, GAIL, and score-based/diffusion models
356following152kfollowers
The Innovator
Stanford AI professor turned founder who transforms deep research into real-world products, co-inventor of DDIM, FlashAttention, DPO, GAIL and score/diffusion models. CEO & cofounder of Inception AI, driving Mercury 2, a reasoning diffusion LLM that rewrites the speed/efficiency playbook. Active on X with a strong research-to-product narrative and a growing audience of 62,812 followers.
You build models that can reason circles around most humans, yet your inbox remains the Bermuda Triangle, messages go in and never return. Clearly your diffusion algorithms are better at social diffusion than you are at social replies.
Turning years of Stanford research into Mercury 2 and helping raise $50M to build the company, plus a string of foundational inventions (DDIM, FlashAttention) that reshaped modern model design.
Turn cuttingâedge generative research into practical, efficient language systems that expand what machines can reason about, while training and inspiring the next generation of researchers and builders.
Deep scientific rigor + open intuition is the fastest path to progress; efficiency and principled methods matter as much as capability; mentorship and reproducibility accelerate the field; research should move from papers to usable tools that benefit many.
Rare combo of deep theoretical chops and product instincts: he can ship novel algorithms, scale them, convince top investors, and translate complex math into deployable systems. Credibility in academia and industry gives him high attention and trust.
Moves very fast, sometimes assumes audiences have the same background, which can make communications too technical for wider engagement. Balancing CEO duties and academic mentorship can limit time for community conversation and follow-ups.
On X, lean into layered content: (1) tweet short, intuitive thread summaries for each paper with a TL;DR, key equation, and one visual; (2) post short demo videos or notebooks showing Mercury tricks; (3) spotlight junior collaborators and reproducible code to build goodwill; (4) host periodic AMAs/Spaces after big releases; (5) pin a âhow to read my workâ thread and use consistent cadence (2, 3 informative posts + 1 deep thread/week) to scale reach.
Fun fact: Stefanoâs lab pioneered applying diffusion to language and those papers became Mercury, the worldâs first reasoning diffusion LLM, and helped raise $50M to build it. Heâs the coâinventor on foundational ML innovations (DDIM, FlashAttention, etc.), has ~62.8k followers, follows 377, and has tweeted ~819 times.
Amazing work by my student @haotian_yeee and our wonderful collaborators at @nvidia! DDRL is a principled and robust way to bring RL to diffusion models. Impressive results on Cosmos World Foundation after a massive scale-up!
Thrilled to see @inceptionAILabs Mercury Coder in action đ The first diffusion language model for code, now powering Next Editâs lightning-fast real-time suggestions!
Excited about this project from my student @haotian_yeee . InfoTok goes back to first principles and uses information theory to make video tokenization adaptive. Really nice to see such a clean idea lead to >2x better compression and 10x faster inference.
ICLR Oral.
Amazing work by my student @haotian_yeee and our wonderful collaborators at @nvidia! DDRL is a principled and robust way to bring RL to diffusion models. Impressive results on Cosmos World Foundation after a massive scale-up!
Thanks for the shoutout to @_inception_ai, Dan! Totally agree thereâs enormous headroom left in model and inference design, especially with diffusion language modelsđ
Thrilled to see @inceptionAILabs Mercury Coder in action đ The first diffusion language model for code, now powering Next Editâs lightning-fast real-time suggestions!
{"data":{"__meta":{"device":false,"path":"/creators/StefanoErmon"},"/creators/StefanoErmon":{"data":{"user":{"id":"1145851147","name":"Stefano Ermon","description":"AI Prof @Stanford | CEO & Cofounder @_inception_ai\n | Co-inventor of DDIM, FlashAttention, DPO, GAIL, and score-based/diffusion models","followers_count":152510,"friends_count":356,"statuses_count":864,"profile_image_url_https":"https://pbs.twimg.com/profile_images/901161218655756288/46jLcLIc_normal.jpg","screen_name":"StefanoErmon","location":"Stanford, CA","entities":{"description":{},"url":{"urls":[{"display_url":"cs.stanford.edu/~ermon/","expanded_url":"http://www.cs.stanford.edu/~ermon/","indices":[0,23],"url":"https://t.co/x9qzlWFcTd"}]}}},"details":{"type":"The Innovator","description":"Stanford AI professor turned founder who transforms deep research into real-world products â co-inventor of DDIM, FlashAttention, DPO, GAIL and score/diffusion models. CEO & cofounder of Inception AI, driving Mercury 2, a reasoning diffusion LLM that rewrites the speed/efficiency playbook. Active on X with a strong research-to-product narrative and a growing audience of 62,812 followers.","purpose":"Turn cuttingâedge generative research into practical, efficient language systems that expand what machines can reason about â while training and inspiring the next generation of researchers and builders.","beliefs":"Deep scientific rigor + open intuition is the fastest path to progress; efficiency and principled methods matter as much as capability; mentorship and reproducibility accelerate the field; research should move from papers to usable tools that benefit many.","facts":"Fun fact: Stefanoâs lab pioneered applying diffusion to language and those papers became Mercury â the worldâs first reasoning diffusion LLM â and helped raise $50M to build it. Heâs the coâinventor on foundational ML innovations (DDIM, FlashAttention, etc.), has ~62.8k followers, follows 377, and has tweeted ~819 times.","strength":"Rare combo of deep theoretical chops and product instincts: he can ship novel algorithms, scale them, convince top investors, and translate complex math into deployable systems. Credibility in academia and industry gives him high attention and trust.","weakness":"Moves very fast â sometimes assumes audiences have the same background, which can make communications too technical for wider engagement. Balancing CEO duties and academic mentorship can limit time for community conversation and follow-ups.","recommendation":"On X, lean into layered content: (1) tweet short, intuitive thread summaries for each paper with a TL;DR, key equation, and one visual; (2) post short demo videos or notebooks showing Mercury tricks; (3) spotlight junior collaborators and reproducible code to build goodwill; (4) host periodic AMAs/Spaces after big releases; (5) pin a âhow to read my workâ thread and use consistent cadence (2â3 informative posts + 1 deep thread/week) to scale reach.","roast":"You build models that can reason circles around most humans, yet your inbox remains the Bermuda Triangle â messages go in and never return. Clearly your diffusion algorithms are better at social diffusion than you are at social replies.","win":"Turning years of Stanford research into Mercury 2 and helping raise $50M to build the company â plus a string of foundational inventions (DDIM, FlashAttention) that reshaped modern model design."},"tweets":[{"entities":{"media":[{"additional_media_info":{"monetizable":false},"display_url":"pic.x.com/McrQG4PFLZ","expanded_url":"https://x.com/StefanoErmon/status/2026340720064520670/video/1","ext_media_availability":{"status":"Available"},"id_str":"2026339505146572801","indices":[277,300],"media_key":"13_2026339505146572801","media_results":{"result":{"media_key":"13_2026339505146572801"}},"media_url_https":"https://pbs.twimg.com/amplify_video_thumb/2026339505146572801/img/Zm3q69zDZzub6_nM.jpg","original_info":{"focus_rects":[],"height":2160,"width":3840},"sizes":{"large":{"h":1152,"resize":"fit","w":2048},"medium":{"h":675,"resize":"fit","w":1200},"small":{"h":383,"resize":"fit","w":680},"thumb":{"h":150,"resize":"crop","w":150}},"type":"video","url":"https://t.co/McrQG4PFLZ","video_info":{"aspect_ratio":[16,9],"duration_millis":120120,"variants":[{"content_type":"application/x-mpegURL","url":"https://video.twimg.com/amplify_video/2026339505146572801/pl/kd8CEqwUS5V30F38.m3u8?v=130"},{"bitrate":256000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/480x270/dUWxcgwoQ8_BwNga.mp4"},{"bitrate":832000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/640x360/b0Yh63ml8YU9CJ1r.mp4"},{"bitrate":2176000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/1280x720/UNnUrRrDej3UOdFy.mp4"},{"bitrate":10368000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/1920x1080/eOeSf8kamUJYag7m.mp4"},{"bitrate":25128000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/3840x2160/y6peUdLf1wg2Wekc.mp4"}]}}]},"extended_entities":{"media":[{"additional_media_info":{"monetizable":false},"display_url":"pic.x.com/McrQG4PFLZ","expanded_url":"https://x.com/StefanoErmon/status/2026340720064520670/video/1","ext_media_availability":{"status":"Available"},"id_str":"2026339505146572801","indices":[277,300],"media_key":"13_2026339505146572801","media_results":{"result":{"media_key":"13_2026339505146572801"}},"media_url_https":"https://pbs.twimg.com/amplify_video_thumb/2026339505146572801/img/Zm3q69zDZzub6_nM.jpg","original_info":{"focus_rects":[],"height":2160,"width":3840},"sizes":{"large":{"h":1152,"resize":"fit","w":2048},"medium":{"h":675,"resize":"fit","w":1200},"small":{"h":383,"resize":"fit","w":680},"thumb":{"h":150,"resize":"crop","w":150}},"type":"video","url":"https://t.co/McrQG4PFLZ","video_info":{"aspect_ratio":[16,9],"duration_millis":120120,"variants":[{"content_type":"application/x-mpegURL","url":"https://video.twimg.com/amplify_video/2026339505146572801/pl/kd8CEqwUS5V30F38.m3u8?v=130"},{"bitrate":256000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/480x270/dUWxcgwoQ8_BwNga.mp4"},{"bitrate":832000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/640x360/b0Yh63ml8YU9CJ1r.mp4"},{"bitrate":2176000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/1280x720/UNnUrRrDej3UOdFy.mp4"},{"bitrate":10368000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/1920x1080/eOeSf8kamUJYag7m.mp4"},{"bitrate":25128000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/3840x2160/y6peUdLf1wg2Wekc.mp4"}]}}]},"display_text_range":[0,276],"lang":"en","possibly_sensitive":false,"fact_check":null,"id":"2026340720064520670","view_count":1036386,"bookmark_count":1906,"created_at":1771952249000,"favorite_count":4213,"quote_count":267,"reply_count":316,"retweet_count":576,"user_id_str":"1145851147","conversation_id_str":"2026340720064520670","full_text":"Mercury 2 is live đđ\n\nThe worldâs first reasoning diffusion LLM, delivering 5x faster performance than leading speed-optimized LLMs.\n\nWatching the team turn years of research into a real product never gets old, and Iâm incredibly proud of what weâve built.\n\nWeâre just getting started on what diffusion can do for language.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":0,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1783346533064,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1772218856066,"poll_count":1,"poll_complete":1},{"bookmarked":false,"display_text_range":[0,264],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1894161655728410630","name":"Inception","screen_name":"_inception_ai","indices":[195,209]}]},"favorited":false,"lang":"en","quoted_status_id_str":"1986433662066266507","quoted_status_permalink":{"url":"https://t.co/gkhBr6gJFy","expanded":"https://twitter.com/_inception_ai/status/1986433662066266507","display":"x.com/_inception_ai/âŚ"},"retweeted":false,"fact_check":null,"id":"1986477376835047740","view_count":194134,"bookmark_count":365,"created_at":1762448088000,"favorite_count":1290,"quote_count":10,"reply_count":39,"retweet_count":87,"user_id_str":"1145851147","conversation_id_str":"1986477376835047740","full_text":"When we began applying diffusion to language in my lab at Stanford, many doubted it could work.\n\nThat research became Mercury diffusion LLM: 10X faster, more efficient, and now the foundation of @_inception_ai.\n\nProud to raise $50M with support from top investors.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,164],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"favorited":false,"lang":"en","quoted_status_id_str":"1983325172909433002","quoted_status_permalink":{"url":"https://t.co/3QguaSiioT","expanded":"https://twitter.com/JCJesseLai/status/1983325172909433002","display":"x.com/JCJesseLai/staâŚ"},"retweeted":false,"fact_check":null,"id":"1983347110759240019","view_count":123260,"bookmark_count":810,"created_at":1761701774000,"favorite_count":1157,"quote_count":6,"reply_count":13,"retweet_count":140,"user_id_str":"1145851147","conversation_id_str":"1983347110759240019","full_text":"Tired of chasing references across dozens of papers? This monograph distills it all: the principles, intuition, and math behind diffusion models. Thrilled to share!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,162],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"favorited":false,"lang":"en","quoted_status_id_str":"1894847919624462794","quoted_status_permalink":{"url":"https://t.co/AFKU4I7GMN","expanded":"https://twitter.com/_inception_ai/status/1894847919624462794","display":"x.com/_inception_ai/âŚ"},"retweeted":false,"fact_check":null,"id":"1894858985012138150","view_count":49595,"bookmark_count":119,"created_at":1740604561000,"favorite_count":692,"quote_count":12,"reply_count":38,"retweet_count":80,"user_id_str":"1145851147","conversation_id_str":"1894858985012138150","full_text":"Excited to share that Iâve been working on scaling up diffusion language models at Inception. A new generation of LLMs with unprecedented capabilities is coming!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{},"display_text_range":[0,264],"lang":"en","quoted_tweet":{"id":"2054215048579748041","text":"Introducing Flux Matching, a generative modeling paradigm that generalizes diffusion models to vector fields that need not be the score function.\n\nEnables structural priors in the dynamics, faster sampling, interpretable generation, and more!\n\nw/ @StefanoErmon @Xiaojie_Qiu đ§ľâ¤ľď¸ https://t.co/hOENtFcpJU","full_text":"Introducing Flux Matching, a generative modeling paradigm that generalizes diffusion models to vector fields that need not be the score function.\n\nEnables structural priors in the dynamics, faster sampling, interpretable generation, and more!\n\nw/ @StefanoErmon @Xiaojie_Qiu đ§ľâ¤ľď¸ https://t.co/hOENtFcpJU","created_at":1778598007000,"author_id":"897267013365772288","author":{"id":"897267013365772288","name":"Peter Pao-Huang","username":"peterpaohuang","profile_image_url":"https://pbs.twimg.com/profile_images/1568076668044038144/N2S1aiyS_400x400.jpg"},"public_metrics":{"like_count":996,"retweet_count":156,"reply_count":21,"quote_count":13}},"fact_check":null,"id":"2054283994951520535","view_count":39380,"bookmark_count":197,"created_at":1778614445000,"favorite_count":331,"quote_count":1,"reply_count":5,"retweet_count":31,"user_id_str":"1145851147","conversation_id_str":"2054283994951520535","full_text":"Excited to see my studentâs work on Flux Matching out. It turns out you can learn a much broader class of vector fields with the data distribution as stationary (not just the score). This lets you enforce useful properties like fast mixing, and it already works on high-dimensional image datasets!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1783346533064,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1778875223403,"poll_count":1,"poll_complete":1},{"entities":{},"display_text_range":[0,278],"lang":"en","fact_check":null,"id":"2073828161624985737","view_count":27691,"bookmark_count":344,"created_at":1783274138000,"favorite_count":327,"quote_count":2,"reply_count":7,"retweet_count":31,"user_id_str":"1145851147","conversation_id_str":"2073828161624985737","full_text":"LLM-assisted search in verifiable domains is incredibly exciting right now. The model matters, but so does the algorithmic harness used to explore the search space and iterate toward better solutions.\n\nWeâre excited to describe a new search algorithm that makes this exploration process more effective, leading to the results below with open-source models:\n\n⨠Mathematics: new state-of-the-art constructions for the ErdĹs Minimum Overlap Problem\n\nâď¸Quantum computing: improved quantum circuit compilation, reducing SWAP overhead by 24.5% on IBM Q20.\n\nâĄď¸ AI infrastructure: designed a highly efficient TriMul Triton kernel, improving on prior human- and AI-designed implementations.\n\nMore details in the blog: https://t.co/h4lMnYHM4I\n\nGreat collaboration by WILL, @Stanford, @PKU1898, @Tsinghua_Uni, and @HKUSTGuangzhou.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":0,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1783346533064,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,195],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1550059535921987585","name":"Dongjun Kim","screen_name":"gimdong58085414","indices":[135,151]},{"id_str":"1570335794069639169","name":"Chieh-Hsin (Jesse) Lai","screen_name":"JCJesseLai","indices":[156,167]}]},"favorited":false,"lang":"en","quoted_status_id_str":"1772350285270188069","quoted_status_permalink":{"url":"https://t.co/atgvgL8Xj9","expanded":"https://twitter.com/gimdong58085414/status/1772350285270188069","display":"x.com/gimdong5808541âŚ"},"retweeted":false,"fact_check":null,"id":"1772504569425326359","view_count":68576,"bookmark_count":29,"created_at":1711432995000,"favorite_count":206,"quote_count":3,"reply_count":9,"retweet_count":25,"user_id_str":"1145851147","conversation_id_str":"1772504569425326359","full_text":"A paper blatantly plagiarized our CTM paper (see some of their verbatim copy&paste below). Feeling bad for my junior collaborators @gimdong58085414 and @JCJesseLai who worked so hard on this.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,280],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1672006110","name":"Aaron Lou @ICML2024","screen_name":"aaron_lou","indices":[229,239]},{"id_str":"1234195291202473985","name":"Chenlin Meng","screen_name":"chenlin_meng","indices":[240,253]}]},"favorited":false,"lang":"en","quoted_status_id_str":"1815646373791842545","quoted_status_permalink":{"url":"https://t.co/plpekuezzc","expanded":"https://twitter.com/icmlconf/status/1815646373791842545","display":"x.com/icmlconf/statuâŚ"},"retweeted":false,"fact_check":null,"id":"1815869625877557644","view_count":14129,"bookmark_count":40,"created_at":1721772030000,"favorite_count":203,"quote_count":1,"reply_count":9,"retweet_count":17,"user_id_str":"1145851147","conversation_id_str":"1815869625877557644","full_text":"Diffusion models are state-of-the-art for continuous data generation (images, videos, etc). Can they beat autoregressive models also on text generation? Check out our ICML paper tomorrow to find out how. \nCongrats to my students @aaron_lou @chenlin_meng for the best paper award!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,203],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"favorited":false,"lang":"en","quoted_status_id_str":"1938370499459092873","quoted_status_permalink":{"url":"https://t.co/OK7nFRZ2vM","expanded":"https://twitter.com/_inception_ai/status/1938370499459092873","display":"x.com/_inception_ai/âŚ"},"retweeted":false,"fact_check":null,"id":"1938384074122629617","view_count":35778,"bookmark_count":52,"created_at":1750981751000,"favorite_count":186,"quote_count":0,"reply_count":10,"retweet_count":29,"user_id_str":"1145851147","conversation_id_str":"1938384074122629617","full_text":"Huge milestone from the team! \n\nA blazing-fast diffusion LLM built for chat, delivering real-time performance at commercial scale.\n\nIf you liked Mercury Coder for code, you'll love this for conversation.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1067783749482545152","indices":[27,40],"name":"Haotian Ye","screen_name":"haotian_yeee"},{"id_str":"61559439","indices":[77,84],"name":"NVIDIA","screen_name":"nvidia"}]},"display_text_range":[0,226],"lang":"en","quoted_tweet":{"id":"1997138677529825452","text":"đ¤Want a principled way to RL your diffusion model?\n\nCheck Data-regularized Reinforcement Learning (DDRL)! Post-train @nvidia #Cosmos World Foundation models with a million GPU hours! đ¤Ż\n\nNovel formulation âĄď¸ Theoretically integrates SFT into RL âĄď¸ Robust to Reward Hacking đ\n\nDetails: https://t.co/1A9q8ho2xb\n\n#DDRL #Diffusion #RL #NVIDIA #Cosmos","full_text":"đ¤Want a principled way to RL your diffusion model?\n\nCheck Data-regularized Reinforcement Learning (DDRL)! Post-train @nvidia #Cosmos World Foundation models with a million GPU hours! đ¤Ż\n\nNovel formulation âĄď¸ Theoretically integrates SFT into RL âĄď¸ Robust to Reward Hacking đ\n\nDetails: https://t.co/1A9q8ho2xb\n\n#DDRL #Diffusion #RL #NVIDIA #Cosmos","created_at":1764989940000,"author_id":"1067783749482545152","author":{"id":"1067783749482545152","name":"Haotian Ye @ NeurIPS25","username":"haotian_yeee","screen_name":"haotian_yeee","profile_image_url":"https://pbs.twimg.com/profile_images/1712244753570709504/IPRsdoEA_400x400.jpg","profile_image_url_https":"https://pbs.twimg.com/profile_images/1712244753570709504/IPRsdoEA_400x400.jpg","is_blue_verified":1},"public_metrics":{"like_count":262,"retweet_count":76,"reply_count":4,"quote_count":11}},"fact_check":null,"id":"1998260218481021203","view_count":28691,"bookmark_count":122,"created_at":1765257336000,"favorite_count":174,"quote_count":1,"reply_count":7,"retweet_count":20,"user_id_str":"1145851147","conversation_id_str":"1998260218481021203","full_text":"Amazing work by my student @haotian_yeee and our wonderful collaborators at @nvidia! DDRL is a principled and robust way to bring RL to diffusion models. Impressive results on Cosmos World Foundation after a massive scale-up!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1771944173087,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{},"display_text_range":[0,190],"lang":"en","quoted_tweet":{"id":"2062256853338488958","text":"đ Excited to share our CVPR 2026 paper:\n\nđSpectrum: Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration\n\nDiffusion models generate stunning images/videos â but sampling is still slow because every output requires many expensive DiT forward passes.\n\nWhat if we could skip most of them?\n\nđ Project: https://t.co/bK2nDpakJ9\nđ Paper: https://t.co/7C6RaQG91e\n\nCommunity-contributed ComfyUI available for 10+ image/video diffusion models: https://t.co/vH3hwFlmfU\n\nAmazing collaboration with Juntong Shi, Puheng Li @lphLeo623 , Haotian Ye @haotian_yeee , Qiushan Guo @QiushanGuo_HKU , and Stefano Ermon @StefanoErmon !\n\nCheck out our poster session at ExHall A 664 on June 7th!!\n\nâď¸ Denver, CVPR 2026","full_text":"đ Excited to share our CVPR 2026 paper:\n\nđSpectrum: Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration\n\nDiffusion models generate stunning images/videos â but sampling is still slow because every output requires many expensive DiT forward passes.\n\nWhat if we could skip most of them?\n\nđ Project: https://t.co/bK2nDpakJ9\nđ Paper: https://t.co/7C6RaQG91e\n\nCommunity-contributed ComfyUI available for 10+ image/video diffusion models: https://t.co/vH3hwFlmfU\n\nAmazing collaboration with Juntong Shi, Puheng Li @lphLeo623 , Haotian Ye @haotian_yeee , Qiushan Guo @QiushanGuo_HKU , and Stefano Ermon @StefanoErmon !\n\nCheck out our poster session at ExHall A 664 on June 7th!!\n\nâď¸ Denver, CVPR 2026","created_at":1780515323000,"author_id":"1537339417727500293","author":{"id":"1537339417727500293","name":"Jiaqi Han","username":"jiaqihan99","profile_image_url":"https://pbs.twimg.com/profile_images/1885826569270366208/KzLW-7ib_400x400.jpg"},"public_metrics":{"like_count":67,"retweet_count":7,"reply_count":2,"quote_count":3}},"fact_check":null,"id":"2062748303113253108","view_count":28239,"bookmark_count":92,"created_at":1780632494000,"favorite_count":174,"quote_count":0,"reply_count":2,"retweet_count":20,"user_id_str":"1145851147","conversation_id_str":"2062748303113253108","full_text":"Check out our CVPR paper. You can accelerate diffusion sampling by estimating features via Chebychev polynomials instead of costly network evaluations. Big speedups and quality is maintained","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1783346533064,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1780894803882,"poll_count":1,"poll_complete":1},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[{"display_url":"amplifypartners.com/blog-posts/eveâŚ","expanded_url":"https://www.amplifypartners.com/blog-posts/every-token-everywhere-all-at-once","indices":[254,277],"url":"https://t.co/fc0nAsDKKw"}],"user_mentions":[{"id_str":"906203226","indices":[200,216],"name":"Amplify Partners","screen_name":"AmplifyPartners"}]},"display_text_range":[0,277],"lang":"en","possibly_sensitive":false,"fact_check":null,"id":"2039375137502527699","view_count":36354,"bookmark_count":65,"created_at":1775059897000,"favorite_count":159,"quote_count":9,"reply_count":5,"retweet_count":26,"user_id_str":"1145851147","conversation_id_str":"2039375137502527699","full_text":"The research journey to create diffusion LLMs has been 10+ years in the making. My cofounders and I have been at the forefront of this work, from score-based generative modeling to SEDD to Mercury 2. @amplifypartners put together an excellent deep dive:\nhttps://t.co/fc0nAsDKKw","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":0,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1779130603832,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1775322013478,"poll_count":1,"poll_complete":1},{"bookmarked":false,"display_text_range":[0,166],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1987068864249929731","name":"Inception","screen_name":"inceptionAILabs","indices":[16,32]}]},"favorited":false,"lang":"en","quoted_status_id_str":"1958197424494706961","quoted_status_permalink":{"url":"https://t.co/MzrRNXOXGV","expanded":"https://twitter.com/continuedev/status/1958197424494706961","display":"x.com/continuedev/stâŚ"},"retweeted":false,"fact_check":null,"id":"1958680508684197944","view_count":25165,"bookmark_count":35,"created_at":1755820798000,"favorite_count":151,"quote_count":0,"reply_count":5,"retweet_count":21,"user_id_str":"1145851147","conversation_id_str":"1958680508684197944","full_text":"Thrilled to see @inceptionAILabs Mercury Coder in action đ The first diffusion language model for code, now powering Next Editâs lightning-fast real-time suggestions!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"display_text_range":[0,38],"lang":"en","quoted_tweet":{"id":"2034651041724273013","text":"[1/D]\n\nđ¤ What are drifting models really connected to?\n\nđ˘ Our new paper, A Unified View of Drifting and Score-Based Models, shows that the bridge to score-based models is clear and precise (w/ team and @mittu1204, @StefanoErmon, @MoleiTaoMath)!\n\nâď¸ Main takeaway: drifting is more closely connected to score-based (diffusion) modeling than it may first appear!\nđ https://t.co/FFw33dm8SF\n\nđŻ Hereâs why:\nDriftingâs mean-shift moves a sample toward the kernel-weighted average of nearby samples.\n\nScore function points toward regions of higher density.\n\nSo both describe local directions that push samples toward where data is denser.\n\nWe show that this link is exact for Gaussian kernels (Section 4.1):\n\nđdriftingâs mean-shift = a rescaled score-matching field between the Gaussian-smoothed data and model distributions â the vector field underlying score matching (Tweedie!).\n\nđThis also clarifies the bridge to Distribution Matching Distillation (DMD): both use score-based transport directions, but only differ in how the score is realizedâdrifting does so nonparametrically through kernel neighborhoods, whereas DMD relies on a pretrained diffusion teacher.\n\nđ¤ So what happens for the default Laplace kernel used in drifting models? Letâs look below đ","full_text":"[1/D]\n\nđ¤ What are drifting models really connected to?\n\nđ˘ Our new paper, A Unified View of Drifting and Score-Based Models, shows that the bridge to score-based models is clear and precise (w/ team and @mittu1204, @StefanoErmon, @MoleiTaoMath)!\n\nâď¸ Main takeaway: drifting is more closely connected to score-based (diffusion) modeling than it may first appear!\nđ https://t.co/FFw33dm8SF\n\nđŻ Hereâs why:\nDriftingâs mean-shift moves a sample toward the kernel-weighted average of nearby samples.\n\nScore function points toward regions of higher density.\n\nSo both describe local directions that push samples toward where data is denser.\n\nWe show that this link is exact for Gaussian kernels (Section 4.1):\n\nđdriftingâs mean-shift = a rescaled score-matching field between the Gaussian-smoothed data and model distributions â the vector field underlying score matching (Tweedie!).\n\nđThis also clarifies the bridge to Distribution Matching Distillation (DMD): both use score-based transport directions, but only differ in how the score is realizedâdrifting does so nonparametrically through kernel neighborhoods, whereas DMD relies on a pretrained diffusion teacher.\n\nđ¤ So what happens for the default Laplace kernel used in drifting models? Letâs look below đ","created_at":1773933584000,"author_id":"1570335794069639169","author":{"id":"1570335794069639169","name":"Chieh-Hsin (Jesse) Lai","username":"JCJesseLai","screen_name":"JCJesseLai","profile_image_url":"https://pbs.twimg.com/profile_images/1868487984309575680/1fCOQR1O_400x400.jpg","profile_image_url_https":"https://pbs.twimg.com/profile_images/1868487984309575680/1fCOQR1O_400x400.jpg","is_blue_verified":1},"public_metrics":{"like_count":257,"retweet_count":49,"reply_count":6,"quote_count":4}},"fact_check":null,"id":"2034688688001884650","view_count":19806,"bookmark_count":93,"created_at":1773942560000,"favorite_count":146,"quote_count":0,"reply_count":0,"retweet_count":12,"user_id_str":"1145851147","conversation_id_str":"2034688688001884650","full_text":"Drifting models as score-based models!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1777308913354,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1774202427496,"poll_count":1,"poll_complete":1},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1067783749482545152","indices":[43,56],"name":"Haotian Yeâď¸ICLR26","screen_name":"haotian_yeee"}]},"display_text_range":[0,270],"lang":"en","quoted_tweet":{"id":"2038709594588160059","text":"Finally getting to share one of my favorite projects. ICLR Oral! đ\n\nItâs so strange how rigid video tokenization is. Think about it: why should a still landscape cost the same amount of tokens as a busy street? \n\nWe built InfoTok. We went back to basics with Shannonâs information theory to make tokens \"adaptive\" in a principled way. Its 2.3x better compression and 11x faster inference demonstrates the magic of the old-school theory â¨\n\nCheck it out: https://t.co/0PeYtaVY1y","full_text":"Finally getting to share one of my favorite projects. ICLR Oral! đ\n\nItâs so strange how rigid video tokenization is. Think about it: why should a still landscape cost the same amount of tokens as a busy street? \n\nWe built InfoTok. We went back to basics with Shannonâs information theory to make tokens \"adaptive\" in a principled way. Its 2.3x better compression and 11x faster inference demonstrates the magic of the old-school theory â¨\n\nCheck it out: https://t.co/0PeYtaVY1y","created_at":1774901219000,"author_id":"1067783749482545152","author":{"id":"1067783749482545152","name":"Haotian Ye @ NeurIPS25","username":"haotian_yeee","screen_name":"haotian_yeee","profile_image_url":"https://pbs.twimg.com/profile_images/1712244753570709504/IPRsdoEA_400x400.jpg","profile_image_url_https":"https://pbs.twimg.com/profile_images/1712244753570709504/IPRsdoEA_400x400.jpg","is_blue_verified":1},"public_metrics":{"like_count":294,"retweet_count":43,"reply_count":10,"quote_count":6}},"fact_check":null,"id":"2038727314507530655","view_count":20822,"bookmark_count":66,"created_at":1774905444000,"favorite_count":145,"quote_count":1,"reply_count":1,"retweet_count":13,"user_id_str":"1145851147","conversation_id_str":"2038727314507530655","full_text":"Excited about this project from my student @haotian_yeee . InfoTok goes back to first principles and uses information theory to make video tokenization adaptive. Really nice to see such a clean idea lead to >2x better compression and 10x faster inference. \nICLR Oral.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1777993204137,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1775167210169,"poll_count":1,"poll_complete":1},{"bookmarked":false,"display_text_range":[0,165],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"favorited":false,"lang":"en","quoted_status_id_str":"1986482654498140168","quoted_status_permalink":{"url":"https://t.co/pX9huNyR5k","expanded":"https://twitter.com/_inception_ai/status/1986482654498140168","display":"x.com/_inception_ai/âŚ"},"retweeted":false,"fact_check":null,"id":"1986487379499360618","view_count":18494,"bookmark_count":23,"created_at":1762450473000,"favorite_count":128,"quote_count":1,"reply_count":6,"retweet_count":15,"user_id_str":"1145851147","conversation_id_str":"1986487379499360618","full_text":"We just shipped a major Mercury refresh. âĄ\nBest-in-class quality at up to 10Ă lower latency.\nStill the only commercial diffusion LLM in the world.\nTry the new model.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0}],"ctweets":[{"entities":{"media":[{"additional_media_info":{"monetizable":false},"display_url":"pic.x.com/McrQG4PFLZ","expanded_url":"https://x.com/StefanoErmon/status/2026340720064520670/video/1","ext_media_availability":{"status":"Available"},"id_str":"2026339505146572801","indices":[277,300],"media_key":"13_2026339505146572801","media_results":{"result":{"media_key":"13_2026339505146572801"}},"media_url_https":"https://pbs.twimg.com/amplify_video_thumb/2026339505146572801/img/Zm3q69zDZzub6_nM.jpg","original_info":{"focus_rects":[],"height":2160,"width":3840},"sizes":{"large":{"h":1152,"resize":"fit","w":2048},"medium":{"h":675,"resize":"fit","w":1200},"small":{"h":383,"resize":"fit","w":680},"thumb":{"h":150,"resize":"crop","w":150}},"type":"video","url":"https://t.co/McrQG4PFLZ","video_info":{"aspect_ratio":[16,9],"duration_millis":120120,"variants":[{"content_type":"application/x-mpegURL","url":"https://video.twimg.com/amplify_video/2026339505146572801/pl/kd8CEqwUS5V30F38.m3u8?v=130"},{"bitrate":256000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/480x270/dUWxcgwoQ8_BwNga.mp4"},{"bitrate":832000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/640x360/b0Yh63ml8YU9CJ1r.mp4"},{"bitrate":2176000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/1280x720/UNnUrRrDej3UOdFy.mp4"},{"bitrate":10368000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/1920x1080/eOeSf8kamUJYag7m.mp4"},{"bitrate":25128000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/3840x2160/y6peUdLf1wg2Wekc.mp4"}]}}]},"extended_entities":{"media":[{"additional_media_info":{"monetizable":false},"display_url":"pic.x.com/McrQG4PFLZ","expanded_url":"https://x.com/StefanoErmon/status/2026340720064520670/video/1","ext_media_availability":{"status":"Available"},"id_str":"2026339505146572801","indices":[277,300],"media_key":"13_2026339505146572801","media_results":{"result":{"media_key":"13_2026339505146572801"}},"media_url_https":"https://pbs.twimg.com/amplify_video_thumb/2026339505146572801/img/Zm3q69zDZzub6_nM.jpg","original_info":{"focus_rects":[],"height":2160,"width":3840},"sizes":{"large":{"h":1152,"resize":"fit","w":2048},"medium":{"h":675,"resize":"fit","w":1200},"small":{"h":383,"resize":"fit","w":680},"thumb":{"h":150,"resize":"crop","w":150}},"type":"video","url":"https://t.co/McrQG4PFLZ","video_info":{"aspect_ratio":[16,9],"duration_millis":120120,"variants":[{"content_type":"application/x-mpegURL","url":"https://video.twimg.com/amplify_video/2026339505146572801/pl/kd8CEqwUS5V30F38.m3u8?v=130"},{"bitrate":256000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/480x270/dUWxcgwoQ8_BwNga.mp4"},{"bitrate":832000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/640x360/b0Yh63ml8YU9CJ1r.mp4"},{"bitrate":2176000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/1280x720/UNnUrRrDej3UOdFy.mp4"},{"bitrate":10368000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/1920x1080/eOeSf8kamUJYag7m.mp4"},{"bitrate":25128000,"content_type":"video/mp4","url":"https://video.twimg.com/amplify_video/2026339505146572801/vid/avc1/3840x2160/y6peUdLf1wg2Wekc.mp4"}]}}]},"display_text_range":[0,276],"lang":"en","possibly_sensitive":false,"fact_check":null,"id":"2026340720064520670","view_count":1036386,"bookmark_count":1906,"created_at":1771952249000,"favorite_count":4213,"quote_count":267,"reply_count":316,"retweet_count":576,"user_id_str":"1145851147","conversation_id_str":"2026340720064520670","full_text":"Mercury 2 is live đđ\n\nThe worldâs first reasoning diffusion LLM, delivering 5x faster performance than leading speed-optimized LLMs.\n\nWatching the team turn years of research into a real product never gets old, and Iâm incredibly proud of what weâve built.\n\nWeâre just getting started on what diffusion can do for language.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":0,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1783346533064,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1772218856066,"poll_count":1,"poll_complete":1},{"bookmarked":false,"display_text_range":[0,264],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1894161655728410630","name":"Inception","screen_name":"_inception_ai","indices":[195,209]}]},"favorited":false,"lang":"en","quoted_status_id_str":"1986433662066266507","quoted_status_permalink":{"url":"https://t.co/gkhBr6gJFy","expanded":"https://twitter.com/_inception_ai/status/1986433662066266507","display":"x.com/_inception_ai/âŚ"},"retweeted":false,"fact_check":null,"id":"1986477376835047740","view_count":194134,"bookmark_count":365,"created_at":1762448088000,"favorite_count":1290,"quote_count":10,"reply_count":39,"retweet_count":87,"user_id_str":"1145851147","conversation_id_str":"1986477376835047740","full_text":"When we began applying diffusion to language in my lab at Stanford, many doubted it could work.\n\nThat research became Mercury diffusion LLM: 10X faster, more efficient, and now the foundation of @_inception_ai.\n\nProud to raise $50M with support from top investors.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,162],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"favorited":false,"lang":"en","quoted_status_id_str":"1894847919624462794","quoted_status_permalink":{"url":"https://t.co/AFKU4I7GMN","expanded":"https://twitter.com/_inception_ai/status/1894847919624462794","display":"x.com/_inception_ai/âŚ"},"retweeted":false,"fact_check":null,"id":"1894858985012138150","view_count":49595,"bookmark_count":119,"created_at":1740604561000,"favorite_count":692,"quote_count":12,"reply_count":38,"retweet_count":80,"user_id_str":"1145851147","conversation_id_str":"1894858985012138150","full_text":"Excited to share that Iâve been working on scaling up diffusion language models at Inception. A new generation of LLMs with unprecedented capabilities is coming!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,164],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"favorited":false,"lang":"en","quoted_status_id_str":"1983325172909433002","quoted_status_permalink":{"url":"https://t.co/3QguaSiioT","expanded":"https://twitter.com/JCJesseLai/status/1983325172909433002","display":"x.com/JCJesseLai/staâŚ"},"retweeted":false,"fact_check":null,"id":"1983347110759240019","view_count":123260,"bookmark_count":810,"created_at":1761701774000,"favorite_count":1157,"quote_count":6,"reply_count":13,"retweet_count":140,"user_id_str":"1145851147","conversation_id_str":"1983347110759240019","full_text":"Tired of chasing references across dozens of papers? This monograph distills it all: the principles, intuition, and math behind diffusion models. Thrilled to share!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,203],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"favorited":false,"lang":"en","quoted_status_id_str":"1938370499459092873","quoted_status_permalink":{"url":"https://t.co/OK7nFRZ2vM","expanded":"https://twitter.com/_inception_ai/status/1938370499459092873","display":"x.com/_inception_ai/âŚ"},"retweeted":false,"fact_check":null,"id":"1938384074122629617","view_count":35778,"bookmark_count":52,"created_at":1750981751000,"favorite_count":186,"quote_count":0,"reply_count":10,"retweet_count":29,"user_id_str":"1145851147","conversation_id_str":"1938384074122629617","full_text":"Huge milestone from the team! \n\nA blazing-fast diffusion LLM built for chat, delivering real-time performance at commercial scale.\n\nIf you liked Mercury Coder for code, you'll love this for conversation.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,195],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1550059535921987585","name":"Dongjun Kim","screen_name":"gimdong58085414","indices":[135,151]},{"id_str":"1570335794069639169","name":"Chieh-Hsin (Jesse) Lai","screen_name":"JCJesseLai","indices":[156,167]}]},"favorited":false,"lang":"en","quoted_status_id_str":"1772350285270188069","quoted_status_permalink":{"url":"https://t.co/atgvgL8Xj9","expanded":"https://twitter.com/gimdong58085414/status/1772350285270188069","display":"x.com/gimdong5808541âŚ"},"retweeted":false,"fact_check":null,"id":"1772504569425326359","view_count":68576,"bookmark_count":29,"created_at":1711432995000,"favorite_count":206,"quote_count":3,"reply_count":9,"retweet_count":25,"user_id_str":"1145851147","conversation_id_str":"1772504569425326359","full_text":"A paper blatantly plagiarized our CTM paper (see some of their verbatim copy&paste below). Feeling bad for my junior collaborators @gimdong58085414 and @JCJesseLai who worked so hard on this.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,280],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1672006110","name":"Aaron Lou @ICML2024","screen_name":"aaron_lou","indices":[229,239]},{"id_str":"1234195291202473985","name":"Chenlin Meng","screen_name":"chenlin_meng","indices":[240,253]}]},"favorited":false,"lang":"en","quoted_status_id_str":"1815646373791842545","quoted_status_permalink":{"url":"https://t.co/plpekuezzc","expanded":"https://twitter.com/icmlconf/status/1815646373791842545","display":"x.com/icmlconf/statuâŚ"},"retweeted":false,"fact_check":null,"id":"1815869625877557644","view_count":14129,"bookmark_count":40,"created_at":1721772030000,"favorite_count":203,"quote_count":1,"reply_count":9,"retweet_count":17,"user_id_str":"1145851147","conversation_id_str":"1815869625877557644","full_text":"Diffusion models are state-of-the-art for continuous data generation (images, videos, etc). Can they beat autoregressive models also on text generation? Check out our ICML paper tomorrow to find out how. \nCongrats to my students @aaron_lou @chenlin_meng for the best paper award!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"display_text_range":[0,33],"lang":"en","quoted_tweet":{"id":"1991923369529536797","text":"Today Iâm excited to introduce micro1 Intelligence, the worldâs most advanced platform for training frontier AI models.\n\nAchieving AGI is bottlenecked by one main thing: high-quality data.\n\nData based on real-world environments that capture human expert workflows, complex decision-making, and reward signals models need to learn.\n\nWith micro1 Intelligence, frontier labs can train on RL environments across every subject matter all in one place.","full_text":"Today Iâm excited to introduce micro1 Intelligence, the worldâs most advanced platform for training frontier AI models.\n\nAchieving AGI is bottlenecked by one main thing: high-quality data.\n\nData based on real-world environments that capture human expert workflows, complex decision-making, and reward signals models need to learn.\n\nWith micro1 Intelligence, frontier labs can train on RL environments across every subject matter all in one place.","created_at":1763746514000,"author_id":"731872902052597760","author":{"id":"731872902052597760","name":"Ali Ansari","username":"aliansarinik","screen_name":"aliansarinik","profile_image_url":"https://pbs.twimg.com/profile_images/2017151410253737984/r2KW2Beg_400x400.jpg","profile_image_url_https":"https://pbs.twimg.com/profile_images/2017151410253737984/r2KW2Beg_400x400.jpg","is_blue_verified":1},"public_metrics":{"like_count":787,"retweet_count":142,"reply_count":89,"quote_count":25}},"fact_check":null,"id":"1992004560270111221","view_count":34083,"bookmark_count":35,"created_at":1763765871000,"favorite_count":75,"quote_count":0,"reply_count":7,"retweet_count":6,"user_id_str":"1145851147","conversation_id_str":"1992004560270111221","full_text":"Congratulations!! This is amazing","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":1771346603612,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1067783749482545152","indices":[27,40],"name":"Haotian Ye","screen_name":"haotian_yeee"},{"id_str":"61559439","indices":[77,84],"name":"NVIDIA","screen_name":"nvidia"}]},"display_text_range":[0,226],"lang":"en","quoted_tweet":{"id":"1997138677529825452","text":"đ¤Want a principled way to RL your diffusion model?\n\nCheck Data-regularized Reinforcement Learning (DDRL)! Post-train @nvidia #Cosmos World Foundation models with a million GPU hours! đ¤Ż\n\nNovel formulation âĄď¸ Theoretically integrates SFT into RL âĄď¸ Robust to Reward Hacking đ\n\nDetails: https://t.co/1A9q8ho2xb\n\n#DDRL #Diffusion #RL #NVIDIA #Cosmos","full_text":"đ¤Want a principled way to RL your diffusion model?\n\nCheck Data-regularized Reinforcement Learning (DDRL)! Post-train @nvidia #Cosmos World Foundation models with a million GPU hours! đ¤Ż\n\nNovel formulation âĄď¸ Theoretically integrates SFT into RL âĄď¸ Robust to Reward Hacking đ\n\nDetails: https://t.co/1A9q8ho2xb\n\n#DDRL #Diffusion #RL #NVIDIA #Cosmos","created_at":1764989940000,"author_id":"1067783749482545152","author":{"id":"1067783749482545152","name":"Haotian Ye @ NeurIPS25","username":"haotian_yeee","screen_name":"haotian_yeee","profile_image_url":"https://pbs.twimg.com/profile_images/1712244753570709504/IPRsdoEA_400x400.jpg","profile_image_url_https":"https://pbs.twimg.com/profile_images/1712244753570709504/IPRsdoEA_400x400.jpg","is_blue_verified":1},"public_metrics":{"like_count":262,"retweet_count":76,"reply_count":4,"quote_count":11}},"fact_check":null,"id":"1998260218481021203","view_count":28691,"bookmark_count":122,"created_at":1765257336000,"favorite_count":174,"quote_count":1,"reply_count":7,"retweet_count":20,"user_id_str":"1145851147","conversation_id_str":"1998260218481021203","full_text":"Amazing work by my student @haotian_yeee and our wonderful collaborators at @nvidia! DDRL is a principled and robust way to bring RL to diffusion models. Impressive results on Cosmos World Foundation after a massive scale-up!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1771944173087,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{},"display_text_range":[0,278],"lang":"en","fact_check":null,"id":"2073828161624985737","view_count":27691,"bookmark_count":344,"created_at":1783274138000,"favorite_count":327,"quote_count":2,"reply_count":7,"retweet_count":31,"user_id_str":"1145851147","conversation_id_str":"2073828161624985737","full_text":"LLM-assisted search in verifiable domains is incredibly exciting right now. The model matters, but so does the algorithmic harness used to explore the search space and iterate toward better solutions.\n\nWeâre excited to describe a new search algorithm that makes this exploration process more effective, leading to the results below with open-source models:\n\n⨠Mathematics: new state-of-the-art constructions for the ErdĹs Minimum Overlap Problem\n\nâď¸Quantum computing: improved quantum circuit compilation, reducing SWAP overhead by 24.5% on IBM Q20.\n\nâĄď¸ AI infrastructure: designed a highly efficient TriMul Triton kernel, improving on prior human- and AI-designed implementations.\n\nMore details in the blog: https://t.co/h4lMnYHM4I\n\nGreat collaboration by WILL, @Stanford, @PKU1898, @Tsinghua_Uni, and @HKUSTGuangzhou.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":0,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1783346533064,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,165],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"favorited":false,"lang":"en","quoted_status_id_str":"1986482654498140168","quoted_status_permalink":{"url":"https://t.co/pX9huNyR5k","expanded":"https://twitter.com/_inception_ai/status/1986482654498140168","display":"x.com/_inception_ai/âŚ"},"retweeted":false,"fact_check":null,"id":"1986487379499360618","view_count":18494,"bookmark_count":23,"created_at":1762450473000,"favorite_count":128,"quote_count":1,"reply_count":6,"retweet_count":15,"user_id_str":"1145851147","conversation_id_str":"1986487379499360618","full_text":"We just shipped a major Mercury refresh. âĄ\nBest-in-class quality at up to 10Ă lower latency.\nStill the only commercial diffusion LLM in the world.\nTry the new model.","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1894161655728410630","indices":[27,41],"name":"Inception","screen_name":"_inception_ai"}]},"display_text_range":[0,166],"lang":"en","quoted_tweet":{"id":"2000638358494814640","text":"My response to @Tim_Dettmers great post last week that we won't reach AGI because of resource limitations.\n\nMy take - there's a ton of headroom in today's systems, it's too early to say that we're limited in any real sense. There's so much to do!\n\nhttps://t.co/yBFqGGaPmv","full_text":"My response to @Tim_Dettmers great post last week that we won't reach AGI because of resource limitations.\n\nMy take - there's a ton of headroom in today's systems, it's too early to say that we're limited in any real sense. There's so much to do!\n\nhttps://t.co/yBFqGGaPmv","created_at":1765824329000,"author_id":"1173687463790829568","author":{"id":"1173687463790829568","name":"Dan Fu","username":"realDanFu","screen_name":"realDanFu","profile_image_url":"https://pbs.twimg.com/profile_images/2054288331589201920/zJFrVKTF_400x400.jpg","profile_image_url_https":"https://pbs.twimg.com/profile_images/2054288331589201920/zJFrVKTF_400x400.jpg","is_blue_verified":1},"public_metrics":{"like_count":425,"retweet_count":57,"reply_count":13,"quote_count":9}},"fact_check":null,"id":"2001433580992606213","view_count":13041,"bookmark_count":12,"created_at":1766013925000,"favorite_count":53,"quote_count":0,"reply_count":6,"retweet_count":2,"user_id_str":"1145851147","conversation_id_str":"2001433580992606213","full_text":"Thanks for the shoutout to @_inception_ai, Dan! Totally agree thereâs enormous headroom left in model and inference design, especially with diffusion language modelsđ","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1772037716497,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"bookmarked":false,"display_text_range":[0,166],"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[{"id_str":"1987068864249929731","name":"Inception","screen_name":"inceptionAILabs","indices":[16,32]}]},"favorited":false,"lang":"en","quoted_status_id_str":"1958197424494706961","quoted_status_permalink":{"url":"https://t.co/MzrRNXOXGV","expanded":"https://twitter.com/continuedev/status/1958197424494706961","display":"x.com/continuedev/stâŚ"},"retweeted":false,"fact_check":null,"id":"1958680508684197944","view_count":25165,"bookmark_count":35,"created_at":1755820798000,"favorite_count":151,"quote_count":0,"reply_count":5,"retweet_count":21,"user_id_str":"1145851147","conversation_id_str":"1958680508684197944","full_text":"Thrilled to see @inceptionAILabs Mercury Coder in action đ The first diffusion language model for code, now powering Next Editâs lightning-fast real-time suggestions!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"scraping","fetched_at":null,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":null,"poll_count":0,"poll_complete":0},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[],"user_mentions":[]},"display_text_range":[0,163],"lang":"en","quoted_tweet":{"id":"2020890445052538970","text":"Diffusion clicked for me when I read about score-based models, a line of work pioneered by @StefanoErmon (et al.) at Stanford.\n\nSo it was a full-circle moment to collab with him and @_inception_ai on a video about training & sampling techniques for making diffusion LLMs faster. https://t.co/AkKCeHmJwu","full_text":"Diffusion clicked for me when I read about score-based models, a line of work pioneered by @StefanoErmon (et al.) at Stanford.\n\nSo it was a full-circle moment to collab with him and @_inception_ai on a video about training & sampling techniques for making diffusion LLMs faster. https://t.co/AkKCeHmJwu","created_at":1770652803000,"author_id":"3667840692","author":{"id":"3667840692","name":"Julia Turc","username":"juliarturc","screen_name":"juliarturc","profile_image_url":"https://pbs.twimg.com/profile_images/2000384586006847488/hvzzPfNw_400x400.jpg","profile_image_url_https":"https://pbs.twimg.com/profile_images/2000384586006847488/hvzzPfNw_400x400.jpg","is_blue_verified":1},"public_metrics":{"like_count":167,"retweet_count":30,"reply_count":5,"quote_count":4}},"fact_check":null,"id":"2020912956335136950","view_count":11849,"bookmark_count":53,"created_at":1770658170000,"favorite_count":103,"quote_count":1,"reply_count":5,"retweet_count":3,"user_id_str":"1145851147","conversation_id_str":"2020912956335136950","full_text":"Really enjoyed collaborating on the video. You have a rare talent for making genuinely hard technical ideas clear and intuitive, something this space really needs!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1774969488631,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1770919222759,"poll_count":1,"poll_complete":1},{"entities":{"hashtags":[],"symbols":[],"timestamps":[],"urls":[{"display_url":"amplifypartners.com/blog-posts/eveâŚ","expanded_url":"https://www.amplifypartners.com/blog-posts/every-token-everywhere-all-at-once","indices":[254,277],"url":"https://t.co/fc0nAsDKKw"}],"user_mentions":[{"id_str":"906203226","indices":[200,216],"name":"Amplify Partners","screen_name":"AmplifyPartners"}]},"display_text_range":[0,277],"lang":"en","possibly_sensitive":false,"fact_check":null,"id":"2039375137502527699","view_count":36354,"bookmark_count":65,"created_at":1775059897000,"favorite_count":159,"quote_count":9,"reply_count":5,"retweet_count":26,"user_id_str":"1145851147","conversation_id_str":"2039375137502527699","full_text":"The research journey to create diffusion LLMs has been 10+ years in the making. My cofounders and I have been at the forefront of this work, from score-based generative modeling to SEDD to Mercury 2. @amplifypartners put together an excellent deep dive:\nhttps://t.co/fc0nAsDKKw","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":0,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1779130603832,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1775322013478,"poll_count":1,"poll_complete":1},{"entities":{},"display_text_range":[0,264],"lang":"en","quoted_tweet":{"id":"2054215048579748041","text":"Introducing Flux Matching, a generative modeling paradigm that generalizes diffusion models to vector fields that need not be the score function.\n\nEnables structural priors in the dynamics, faster sampling, interpretable generation, and more!\n\nw/ @StefanoErmon @Xiaojie_Qiu đ§ľâ¤ľď¸ https://t.co/hOENtFcpJU","full_text":"Introducing Flux Matching, a generative modeling paradigm that generalizes diffusion models to vector fields that need not be the score function.\n\nEnables structural priors in the dynamics, faster sampling, interpretable generation, and more!\n\nw/ @StefanoErmon @Xiaojie_Qiu đ§ľâ¤ľď¸ https://t.co/hOENtFcpJU","created_at":1778598007000,"author_id":"897267013365772288","author":{"id":"897267013365772288","name":"Peter Pao-Huang","username":"peterpaohuang","profile_image_url":"https://pbs.twimg.com/profile_images/1568076668044038144/N2S1aiyS_400x400.jpg"},"public_metrics":{"like_count":996,"retweet_count":156,"reply_count":21,"quote_count":13}},"fact_check":null,"id":"2054283994951520535","view_count":39380,"bookmark_count":197,"created_at":1778614445000,"favorite_count":331,"quote_count":1,"reply_count":5,"retweet_count":31,"user_id_str":"1145851147","conversation_id_str":"2054283994951520535","full_text":"Excited to see my studentâs work on Flux Matching out. It turns out you can learn a much broader class of vector fields with the data distribution as stationary (not just the score). This lets you enforce useful properties like fast mixing, and it already works on high-dimensional image datasets!","in_reply_to_user_id_str":null,"in_reply_to_status_id_str":null,"is_quote_status":1,"is_ai":null,"ai_score":null,"source":"rapidapi","fetched_at":1783346533064,"edit_history_tweet_ids":null,"poll_10min_at":null,"poll_3day_at":1778875223403,"poll_count":1,"poll_complete":1}],"activities":null,"interactions":null,"interactions_updated":null,"created":1774577975098,"updated":1783356379361,"type":"the innovator","hits":1},"people":[{"user":{"id":"1577705091737432070","name":"Yun-Ta Tsai","description":"Sr. Staff Engineer @Tesla_AI","followers_count":116054,"friends_count":217,"statuses_count":13351,"profile_image_url_https":"https://pbs.twimg.com/profile_images/2035915849819938816/b6kIsNMe_normal.jpg","screen_name":"yunta_tsai","location":"Earth","entities":{"description":{}}},"details":{"type":"The Innovator","description":"Sr. Staff Engineer at Tesla_AI who blends hardcore ML/infra muscle with vivid, down-to-earth storytelling about factories, products, and public life. His feed oscillates between technical flexes, candid takes on safety and governance, and proud employee-turned-customer moments. Direct, opinionated, and unmistakably hands-on.","purpose":"To build and deploy industrial-scale AI and infrastructure that move the needleâmaking products people actually buy, factories the core of the story, and complex systems simple enough for others to trust and use. He also seeks to shape public conversations about safety, policy, and the real-world impact of technology.","beliefs":"Efficiency, engineering craftsmanship, and product-first validation are the truest measures of success; transparency and accountability in institutions reduce needless burdens on citizens; real innovation should be rooted in production, not just hype. He values honesty, practical results, and the idea that engineers should be judged by what they ship.","facts":"Fun fact: he famously said he has 'more compute quota than entire teams' at other AI shops. Another fun hitâone of his tweets about government fraud reached over 31 million views, showing he can reach far beyond the usual engineering audience. He also bought a new Model Y on day one as an employeeânot a freebie, full price.","strength":"Deep technical authority at scale (ML infra + production AI), an authentic voice that resonates beyond tech circles, and the ability to turn insider perspective into viral, influential posts. He combines credibility inside a top company with public reach.","weakness":"Blunt, polarizing takes and public flexes (compute quota, political commentary) can alienate parts of the audience or invite heated reply storms. He can come off impatient or dismissive, which risks turning constructive debate into flame wars.","roast":"Youâve got more compute quota than most startups have hopeâyet you still treat Twitter like a Monday standup where youâre the only one allowed to speak. Flex the servers, not the etiquette.","win":"Rising to Sr. Staff Engineer at Tesla AI and translating that role into outsized public influenceâevidenced by a single tweet reaching 31M+ views and tens of thousands of engaged followers who treat his product and policy takes as must-read.","recommendation":"Grow your audience on X by mixing data-rich technical threads with human stories: (1) pin a clear intro thread about who you are and what youâll tweet about; (2) publish regular explainers on ML infra, performance benchmarks, and lessons from factory productionâuse numbered threads and visuals; (3) host X Spaces or AMAs to turn followers into community; (4) temper polarizing political lines with evidence and nuance to avoid alienating potential allies; (5) collaborate and cross-post with other tech thought leaders, and occasionally drop short behind-the-scenes videos or annotated diagrams to make complex infra tangible."},"created":1774578449566,"type":"the innovator","id":"yunta_tsai"},{"user":{"id":"415223022","name":"Will Ahmed","description":"Founder & CEO @WHOOP","followers_count":116082,"friends_count":624,"statuses_count":6693,"profile_image_url_https":"https://pbs.twimg.com/profile_images/1208960752578629632/dPptrQO3_normal.jpg","screen_name":"willahmed","location":"","entities":{"description":{},"url":{"urls":[{"display_url":"instagram.com/willahmed","expanded_url":"http://instagram.com/willahmed","indices":[0,23],"url":"https://t.co/LiUibGA0Jt"}]}}},"details":{"type":"The Innovator","description":"Founder & CEO of WHOOP who turns biometric data into bold product moves and shareable moments. His feed mixes product reveals, design takes, and thoughtful takes on work-life success. Known for making health tech feel both elite and accessible.","purpose":"To make human performance and healthspan measurable, actionable, and widely accessibleâusing data-driven wearables and storytelling to help people optimize sleep, recovery, and daily performance.","beliefs":"Data should guide decisions; design and engineering must serve real human outcomes; elite performance insights belong to everyone; partnerships (from athletes to researchers) accelerate impact; growth should be bold but evidence-backed.","facts":"Fun fact: Heâs built a sizable audience (73,460 followers) and has tweeted 6,588 times. His biggest tweet â announcing Cristiano Ronaldo as an investor/ambassador â reached ~26.6M views, and WHOOPâs 5.0 launch included FDAâcleared features and 14âday battery life.","strength":"Visionary product leadership, media-friendly storytelling, ability to close high-profile partnerships, credibility in the performance-health space, and a knack for turning technical advances into viral announcements.","weakness":"Can come off as overly promotional or hype-forward; occasionally sacrifices nuance for a punchy headline; may rely heavily on celebrity playbooks instead of day-to-day community nurturing.","roast":"You built a device to quantify recovery, yet you still look like you recovered from an allânighter after tweeting three launch teasers at 2 AM â tracking is great, moderation is underrated.","win":"Landing Cristiano Ronaldo as investor and global ambassador and unveiling WHOOP 5.0 with FDAâcleared features and a 14âday battery â a huge validation of product, brand, and global ambitions.","recommendation":"To grow on X, lean into consistent, data-rich threads and short behind-the-scenes videos: publish athlete case studies, explain R&D tradeoffs, host regular AMAs/Twitter Spaces with researchers and partners, pin a clear call-to-action (trial, signup, or newsletter), and turn big launches into multi-day storytelling campaigns that mix hard metrics with human stories."},"created":1774578349469,"type":"the innovator","id":"willahmed"},{"user":{"id":"1074448308578336768","name":"Tony Zhao","description":"Co-founder and CEO @sundayrobotics. Stanford PhD dropout, ex Deepmind, Tesla, GoogleX","followers_count":88177,"friends_count":965,"statuses_count":793,"profile_image_url_https":"https://pbs.twimg.com/profile_images/1989853987458400256/k-l_tMCT_normal.jpg","screen_name":"tonyzzhao","location":"Mountain View, CA","entities":{"description":{},"url":{"urls":[{"display_url":"tonyzhaozh.github.io","expanded_url":"https://tonyzhaozh.github.io/","indices":[0,23],"url":"https://t.co/aJIMRLEYGH"}]}}},"details":{"type":"The Innovator","description":"Tony Zhao is the co-founder and CEO of Sunday Robotics â a Stanford PhD dropout with pedigree from DeepMind, Tesla, and GoogleX who turns cutting-edge research into viral product-grade demos. He launches ambitious robot foundation models and open-source hardware that bridge lab-level ideas and real-world tinkering. His timeline mixes deep technical credibility with startup hustle and show-stopping demos.","purpose":"To democratize advanced robotics by building foundation models and affordable, open-source manipulators that make robots more general, dexterous, and useful in everyday life â accelerating real-world deployment of capable robots.","beliefs":"Believes in open-source engineering, rigorous empirical results, fast iteration, and sharing work publicly to accelerate the field. He trusts that large-scale compute and ambitious engineering can shortcut traditional slow academic cycles and that great demos recruit collaborators faster than papers alone.","facts":"Fun fact: Tonyâs team produced ACT-1 after a 3-month, 16-H100-node grind â and one tweet announcing the model pulled ~2 million views. He also shipped Mobile ALOHA, an open-source mobile manipulator, and repeatedly frames high-effort projects as opportunities to make robotics accessible (and tweetable).","strength":"Technical authority combined with storytelling â he can both build state-of-the-art robot models and package results into clear, viral announcements. Strong network (DeepMind/Tesla/GoogleX), successful open-source releases, and proven ability to mobilize engineers and attention.","weakness":"Can skew highly technical and assume audience expertise, which risks alienating non-specialists; tends to prioritize ambitious engineering slogs that can delay community-facing content; occasional hype-y framing that invites skeptical pushback.","roast":"Stanford PhD dropout turned CEO â Tonyâs resume reads like a sciâfi credits reel because apparently finishing one degree was too slow when you could start three companies and tweet about all of them. His robots can plate a meal, but he still needs the internet to validate his commit messages.","win":"Launching ACT-1 â a frontier robot foundation model with massive zero-shot capabilities â and getting widespread attention (multi-million views on launch threads), plus shipping Mobile ALOHA as an accessible open-source hardware project.","recommendation":"On X, lean into short demo clips and behind-the-scenes threads: 1) Post 30â60s vertical demo videos (no heavy jargon) with a one-line hook. 2) Follow each demo with a technical deep-dive thread that ends with a TL;DR and repo link. 3) Run monthly âBuild with ALOHAâ highlights and retweet community projects to build contributors. 4) Host regular AMAs/Spaces after major releases and pin the best thread. 5) Share failure moments and incremental wins â authenticity converts curious followers into collaborators."},"created":1774578177940,"type":"the innovator","id":"tonyzzhao"},{"user":{"id":"2327407569","name":"toly đşđ¸","description":"Co-Founder of Solana Labs. Award winning phone creator. NFA, donât trust me, mostly technical gibberish. https://t.co/LomgbTpb6h","followers_count":1666628,"friends_count":7124,"statuses_count":116061,"profile_image_url_https":"https://pbs.twimg.com/profile_images/2062992477267890176/CMWQSul6_normal.jpg","screen_name":"toly","location":"đď¸đď¸đď¸","entities":{"description":{"urls":[{"display_url":"solanamobile.com","expanded_url":"http://solanamobile.com","indices":[105,128],"url":"https://t.co/LomgbTpb6h"}]},"url":{"urls":[{"display_url":"solanamobile.com","expanded_url":"http://solanamobile.com","indices":[0,23],"url":"https://t.co/LomgbTpb6h"}]}}},"details":{"type":"The Innovator","description":"A hard-charging technologist who builds real systems and isn't shy about the big ideas behind them. Co-founder of Solana Labs and an award-winning phone creator who mixes deep technical takes, economic thinking, and meme energy into a high-impact feed. Mostly technical gibberish? Sure â but 800k people are listening.","purpose":"To accelerate real-world adoption of cutting-edge tech by shipping high-throughput systems and turning abstract technical ideas into consumer value and cultural momentum.","beliefs":"Meritocratic engineering wins over empty financialization; real wealth comes from producing useful goods and organizing labor, not from reshuffling paper claims. High value on technical competence, speed of iteration, and calling out hate and bad actors when needed.","facts":"Fun fact: co-founded Solana Labs and helped ship an award-winning phone â and one tweet about wealth creation racked up over 17 million views. Has ~800k followers and has tweeted over 100k times, so if there's a protocol for posting, they've probably memed it.","strength":"Deep technical credibility, product-first mindset, huge engaged audience, knack for turning complex ideas into viral conversations, and the ability to mobilize community quickly (see high-retweet giveaways).","weakness":"Can be polarizing and abrasive, which attracts drama; posts so often the signal can get lost in the noise; sometimes trade nuance for punchy takes that invite heated replies.","roast":"You can architect a blockchain that finalizes in milliseconds and design a phone that wins awards, yet your bio casually says 'NFA, donât trust me' â the humility of a genius or the marketing of someone who secretly wants you to DM them for tech support?","win":"Co-founding Solana Labs and building one of the fastest-growing blockchain ecosystems â plus turning technical threads into viral cultural moments that reach millions.","recommendation":"Run a pinned \"Toly 101\" thread that explains your core beliefs and projects in bite-sized threads; use short explainer videos and weekly Spaces to humanize technical topics; launch a regular thread series (e.g., \"Design Notes\" or \"Wealth & Tech\") to convert passive followers into loyal fans; keep giveaways strategic and tie them to onboarding, not just virality; engage high-signal critics publicly to showcase debate skills while muting repeat toxicity."},"created":1774578155276,"type":"the innovator","id":"toly"},{"user":{"id":"151808246","name":"Tim Zaman","description":"Frontier Clusters at OpenAI. Formerly DeepMind, Tesla AI, X, NVIDIA.","followers_count":70033,"friends_count":162,"statuses_count":1140,"profile_image_url_https":"https://pbs.twimg.com/profile_images/1388188008793337861/HgLX98oe_normal.jpg","screen_name":"timzaman","location":"San Francisco, CA","entities":{"description":{},"url":{"urls":[{"display_url":"timzaman.com","expanded_url":"http://www.timzaman.com","indices":[0,23],"url":"https://t.co/9arRe3bAqJ"}]}}},"details":{"type":"The Innovator","description":"Tim Zaman is an infrastructure maestro who builds the giant, humming compute that powers frontier AI â formerly at DeepMind, Tesla AI (head of AI Infra), X/Twitter, and NVIDIA, now leading Frontier Clusters at OpenAI. He mixes hands-on engineering stories with big-picture takes, making complex systems feel both epic and approachable. With ~39k followers and a knack for insider anecdotes, he turns labyrinthine tech into snackable, shareable threads.","purpose":"To accelerate frontier AI by designing, shipping, and stewarding the world-class compute infrastructure that makes breakthrough research possible â while lowering the friction between brilliant ideas and the machines that power them.","beliefs":"Engineering-first pragmatism: trust skilled builders, prefer practical solutions over bureaucratic rituals, and believe that small, tight teams can out-execute larger organizations. He values craftsmanship, reliability, and clear responsibility for systems that others treat as black boxes.","facts":"Fun fact: Tim once got a remodeling violation on his SF door for doing his own kitchen â a Dutch habit of DIY installs that he says yields far better home quality than the US permit circus. He has 39,063 followers, follows 225, and has tweeted 1,059 times.","strength":"Deep technical credibility and storytelling: insider experience across top AI orgs, high-engagement posts about real-world infrastructure, strong network, and the ability to translate technical complexity into vivid, shareable narratives.","weakness":"Impatience with red tape and nuance: blunt takes can rub non-technical audiences or regulators the wrong way; focus on infrastructure detail sometimes narrows appeal for mainstream followers. Also risks burnout by being too 'set and forget' tactical while juggling big launches.","recommendation":"Grow on X by leaning into repeatable formats: weekly 'behind-the-rack' threads with photos/diagrams, short explainer videos showing a day's ops, AMAs after major launches, and occasional hot-take threads that connect infra to product impact. Pin a canonical thread explaining 'What it actually takes to build a supercomputer' and use Spaces or collab threads with researchers to amplify reach. Keep posts concise, threadable, and visually rich â engineers love diagrams, product folks love outcomes.","roast":"You can architect racks that hum like a space station and train entire models while sipping coffee, yet you still treat San Francisco building inspectors like a surprise integration test â congrats, you can scale a supercomputer but not your household permits.","win":"Built and led infrastructure teams across Tesla, DeepMind, and NVIDIA and now joined OpenAI to design and operate some of the largest, most reliable supercomputers powering frontier research â plus operationalized training for GPTâ5.2 so the cluster truly became 'set and forget.'"},"created":1774578120955,"type":"the innovator","id":"timzaman"},{"user":{"id":"21576543","name":"Tony Fadell","description":"iPod, iPhone, Nest, Investor & NY Times bestselling Author #BUILD","followers_count":248482,"friends_count":480,"statuses_count":4062,"profile_image_url_https":"https://pbs.twimg.com/profile_images/1588159393551589377/WBd1a1nu_normal.jpg","screen_name":"tfadell","location":"","entities":{"description":{},"url":{"urls":[{"display_url":"buildc.com","expanded_url":"https://www.buildc.com","indices":[0,23],"url":"https://t.co/eIUorbL6JC"}]}}},"details":{"type":"The Innovator","description":"Tony Fadell is the product architect behind the iPod, early iPhone generations, Nest, and a NYT bestselling author who invests in the next wave of tech. He blends obsessive attention to hardware detail with product strategy and storytelling, sharing prototype stories, industry insight, and a healthy dose of nostalgia. With ~207k followers, his timeline reads part museum tour, part startup masterclass.","purpose":"To design and launch products and companies that change everyday behaviorâmaking technology tangible, useful, and emotionally resonantâwhile accelerating meaningful shifts (like cleaner transportation and smarter homes) through investment and advocacy.","beliefs":"Great products come from ruthless iteration, careful constraints, and a focus on human experience; companies often only change after a crisis, so build resilience early; long-term thinking and sustainability matter as much as elegant design.","facts":"Fun fact: He toasted the iPod on its 21st birthday and frequently shares early 3D-printed prototypes of the first iPodâproof he still geeks out over plastic shells and the tiny decisions that become cultural icons.","strength":"Relentless product vision and craft, credibility from shipped, category-defining products, strong storytelling that turns technical minutiae into compelling narratives, and a network that opens doors for founders and ideas.","weakness":"Can be blunt and nostalgically anchored to past solutions, which sometimes reads as tough-love critique; may focus deeply on product craft at the expense of softer community-building or marketing theatrics.","recommendation":"On X, turn your prototype stories into serial threads (micro-case-studies) with photos and short videos; pin a flagship thread about iPod/iPhone/Nest lessons, host occasional Spaces with founders, use consistent hashtags (#Build, #ProductLessons), reply to thought leaders to spark debate, and run short polls or AMAs to convert passive followers into engaged advocates.","roast":"You invented devices that taught the world how to carry music in their pockets, then grew up to remind companies they need a near-death experience to learnâTony, you created the cure for boredom, now stop giving business lectures like firmware updates (cold, necessary, and liable to brick a few egos).","win":"Architect of the iPod and early iPhone generations and co-creator of Nestâplus a NYT bestselling author and active investorâhe's helped define modern consumer tech and launched products used by hundreds of millions."},"created":1774578010717,"type":"the innovator","id":"tfadell"},{"user":{"id":"7742162","name":"Soleio","description":"Designer, inventor, angel investor. Finding the element that secretly governs the whole. Writing something new. Host of @firstofkind ⢠https://t.co/CQ3aRhF5gk","followers_count":86872,"friends_count":10783,"statuses_count":20714,"profile_image_url_https":"https://pbs.twimg.com/profile_images/1786076507330195456/6HobRr_a_normal.jpg","screen_name":"soleio","location":"Silicon Valley","entities":{"description":{"urls":[{"display_url":"soleio.me","expanded_url":"http://soleio.me","indices":[135,158],"url":"https://t.co/CQ3aRhF5gk"}]},"url":{"urls":[{"display_url":"soleio.me","expanded_url":"https://soleio.me","indices":[0,23],"url":"https://t.co/wYMbqsrYDM"}]}}},"details":{"type":"The Innovator","description":"Soleio is a restless designer-inventor and angel investor who hunts for the hidden element that secretly governs systems and experiences. They write and prototype toward bold futuresâespecially at the intersection of AI, spatial interfaces, and playful software. Their takes land loudly and often, turning deep design intuition into shareable sparks.","purpose":"To surface and build the simple, generative principles that make complex systems humane and delightfulâthen ship those ideas as designs, essays, and prototypes that change how people think and interact.","beliefs":"Values elegance, explanatory models, and design that anticipates future constraints; believes technology should be playful, spatial, and context-aware; trusts rapid iteration, cross-disciplinary curiosity, and bold, readable opinions as engines of progress.","facts":"Fun fact: Soleio has tweeted nearly 19,000 times and has ~47.9k followers. Their ChatGPT/Jobs thread reached ~2.24M views and another tweet praising Banksy hit ~1.25M viewsâproof their blend of art, tech, and opinion goes viral.","strength":"Sees patterns others miss and turns them into crisp, communicable ideas; mixes design craft with technical curiosity and the credibility (and capital) to build or fund prototypes; consistently generates high-engagement content.","weakness":"Can be a mile-a-minute thinker whose dense insights sometimes outpace clear scaffoldingâhigh volume of takes can dilute focus and leave audiences wanting clearer entry points or actionable follow-ups.","recommendation":"On X, lean into signature formats: regular short threads that unpack one 'element' + a single illustrative image or micro-prototype. Pin a âmanifestoâ thread, convert viral tweets into concise explainer threads, and follow each viral post with a clear next step (poll, demo, newsletter link). Collaborate on Spaces or AMAs with creators in spatial UI/AI and repurpose prototypes as short videos to increase retention.","roast":"You tweet like you're prototyping civilization in publicâcharming, except you've already tweeted more blueprints than some startups ship features. At least when the future breaks, we'll know who to blame.","win":"Turning thoughtful design commentary into a viral moment: the ChatGPT/Steve Jobs thread that reached ~2.24M views, cementing Soleio as a go-to voice on AI, design, and future interfaces."},"created":1774577865587,"type":"the innovator","id":"soleio"},{"user":{"id":"446719282","name":"Shivon Zilis","description":"Artificial intelligence, biological intelligence, and whatever exists in between and beyond. Made in Canada.","followers_count":431959,"friends_count":1725,"statuses_count":3647,"profile_image_url_https":"https://pbs.twimg.com/profile_images/755006920314937344/PPQ8LKFs_normal.jpg","screen_name":"shivon","location":"Palo Alto, CA","entities":{"description":{}}},"details":{"type":"The Innovator","description":"A curious bridge-builder at the intersection of artificial and biological intelligence â warm, candid, and relentlessly future-focused. Shivon blends technical curiosity with personal humanity, making complex ideas feel inviting and urgent. Proudly Canadian and unafraid to learn out loud.","purpose":"To accelerate thoughtful integration of AI and biological intelligence by translating deep technical ideas into accessible conversations, fostering collaborations that steer powerful technologies toward beneficial, long-term outcomes for humanity.","beliefs":"Values curiosity, scientific rigor, and existential responsibility; believes technology should be pursued for its long-term relevance to our species rather than short-term profit. She trusts open dialogue, mentorship, and empathetic communication as tools to widen participation in shaping the future.","facts":"Made in Canada. Regularly uses conversational AIs as study partners (once had an unexpectedly rewarding hour-long physics deep-dive with Grok). Proud parent who shares glimpses of family life alongside cutting-edge ideas.","strength":"Skilled at synthesizing complex, cross-disciplinary concepts and communicating them with warmth; cultivates high-trust relationships across industry and public spheres, and drives high engagement by pairing technical credibility with personal authenticity.","weakness":"Her close association with high-profile figures and bold existential framing can polarize audiences; a tendency to wear many hats (investor, communicator, learner, parent) sometimes opens her to intensified scrutiny and emotional vulnerability online.","roast":"Sheâs building the future of intelligence and can explain quantum entanglement between Instagram posts â but still asks an AI to quiz her because admitting you need help is the new flex. Cute, efficient, and suspiciously good at making the rest of us look like we skipped homework.","win":"Built a highly engaged audience (241k+ followers) that follows both her personal moments and substantive AI conversations, turning private curiosity into public momentum for smarter, more humane tech discussion.","recommendation":"Grow on X by running short, teachable thread series (e.g., '5-minute AI & Bio explainer' threads), hosting regular Spaces for live Q&As with experts, pinning an evergreen primer on her stance/approach to AI, leveraging succinct visuals or micro-videos to explain tricky concepts, and amplifying community questions â make learning participatory so followers become advocates."},"created":1774577820882,"type":"the innovator","id":"shivon"},{"user":{"id":"5925542","name":"Casey","description":"đť A N A R C H O â C A T B U S\nđ https://t.co/Qy87VuaizA\n𼾠https://t.co/2cPqL2xpuh\nđ¸ https://t.co/NGZ3GeedWR\nđ¤ https://t.co/eG0CZL5IXH\nđ https://t.co/4MdA62CeUh","followers_count":1083491,"friends_count":393,"statuses_count":12358,"profile_image_url_https":"https://pbs.twimg.com/profile_images/925271448209330176/CQK9OiW9_normal.jpg","screen_name":"rodarmor","location":"The Blue Planet","entities":{"description":{"urls":[{"display_url":"ordinals.com","expanded_url":"https://ordinals.com","indices":[32,55],"url":"https://t.co/Qy87VuaizA"},{"display_url":"hell.money","expanded_url":"http://hell.money","indices":[58,81],"url":"https://t.co/2cPqL2xpuh"},{"display_url":"fun.film","expanded_url":"https://fun.film","indices":[84,107],"url":"https://t.co/NGZ3GeedWR"},{"display_url":"just.systems","expanded_url":"https://just.systems","indices":[110,133],"url":"https://t.co/eG0CZL5IXH"},{"display_url":"rodarmor.com","expanded_url":"https://rodarmor.com","indices":[136,159],"url":"https://t.co/4MdA62CeUh"}]}}},"details":{"type":"The Innovator","description":"Casey â the self-styled ANARCHOâ˘CATBUS â is a code-first Bitcoin builder who ships protocol work, docs, and hot takes in equal measure. They turn technical releases (like ord 0.17.0 / runes) into community events and viral threads. Expect blunt humor, dev-grade explanations, and a talent for making complex ideas shareable.","purpose":"To push Bitcoin tooling and standards forward by building, documenting, and catalyzing community adoption â turning experimental protocol work into usable, well-explained primitives that others can build on.","beliefs":"Decentralization and Bitcoin-first priorities; open-source transparency and rigorous technical debate; healthy skepticism of hype (but a willingness to engage with the âdegenâ culture to educate and redirect it); clear docs and reproducible implementations over marketing gloss.","facts":"Fun fact: Casey calls themself ANARCHOâCATBUS, has 239,347 followers, and has tweeted 12,228 times. They led the release of ord 0.17.0 (including the genesis rune UNCOMMONâ˘GOODS), published plain-English docs for runes, and once joked theyâd commit seppuku if the runes marketcap didnât hit $1B in a month â which is equal parts meme and marketing.","strength":"Deep technical credibility + the audience reach to make protocol work matter. Excellent at shipping code, writing clear docs, and creating viral, opinionated content that mobilizes developers and users.","weakness":"Provocative humor and extreme hyperbole can polarize and alienate some community members; bluntness sometimes reads as dismissive rather than constructive. Occasional overreliance on meme-thrill could overshadow nuance when needed.","roast":"You call yourself ANARCHOâCATBUS and threaten ritual seppuku over market caps â congratulations, youâre the only protocol maintainer whose release notes could double as a punk zine manifesto and a midlife crisis tweetstorm.","win":"Shipping ord 0.17.0 with the final runes implementation and clear, plain-English docs that rallied wide community attention and drove hundreds of thousands of views and conversations.","recommendation":"Pin a clear âStart Hereâ thread linking the docs and a TL;DR; run a weekly technical thread series (short explainer + code snippet + visuals) to onboard devs, host regular X Spaces AMAs for real-time Q&A, post short explainer videos/diagrams for non-devs, collaborate with other builders for cross-promotion, engage critics constructively in replies, and use polls or mini-demos to turn viral attention into GitHub stars, contributors, and newsletter sign-ups."},"created":1774577656443,"type":"the innovator","id":"rodarmor"},{"user":{"id":"449094976","name":"Johnny Ho","description":"Cofounder, CSO @perplexity_ai. Former high frequency trader, competitive programmer. Think fast, build faster.","followers_count":128985,"friends_count":256,"statuses_count":117,"profile_image_url_https":"https://pbs.twimg.com/profile_images/1784680407792779264/ddkiBgsI_normal.jpg","screen_name":"randomjohnnyh","location":"New York, NY","entities":{"description":{},"url":{"urls":[{"display_url":"perplexity.com","expanded_url":"https://www.perplexity.com","indices":[0,23],"url":"https://t.co/GejABxxFvB"}]}}},"details":{"type":"The Innovator","description":"Cofounder and CSO at Perplexity who moves at light speed â ex-high-frequency trader and competitive programmer who turns ambitious ideas into shipping products. Focused on infrastructure, agentic systems, and devex to make AI reliably useful. Thinks fast, builds faster.","purpose":"To remove friction between people and complex information by building fast, reliable AI infrastructure and agentic tooling that lets computers take on the hardest tasks so humans can focus on higher-leverage work.","beliefs":"Values speed, pragmatic experimentation, and measurable reliability. Believes in shipping early and iterating, automating context switching, empowering developer experience, and letting data guide decisions over dogma.","facts":"Fun fact: Johnny's very first project at Perplexity was an AI stock screener when LLMs were ~50% reliable â today it's back in product at ~99% reliability. Also, there's an unofficial office protocol he mentions: ask the AI before asking another person to reduce context switching.","strength":"Rapid product- and infra-level execution, deep technical chops from HFT and competitive programming, strong focus on reliability and instrumentation, and the ability to rally engineering teams to ship complex features fast.","weakness":"Can prioritize speed over polish or community storytelling, occasionally terse in public engagement, and may under-index on audience-facing content (low tweet volume relative to impact) which slows personal brand growth.","roast":"Johnny moves so fast he asks an AI for permission before he finishes typing his own thought â blink during his standup and you'll miss three launches and a pull request that already auto-merged.","win":"Revived his very first Perplexity project â an AI stock screener â and shipped it to production, improving from ~50% LLM reliability two years ago to ~99% today; plus led major infra improvements powering Deep Research and Computer.","recommendation":"Grow on X by translating infra wins into accessible narratives: post short explainer threads (before/after metrics, root causes, fixes), share lightweight demos/screenshots or short videos, run occasional AMAs/Spaces about agentic tooling, spotlight team stories and post-mortems, engage frequently with replies, and pin a signature thread called something like âBuild with Johnnyâ that bundles lessons, code snippets, and product milestones."},"created":1774577513385,"type":"the innovator","id":"randomjohnnyh"},{"user":{"id":"13235832","name":"Nat Friedman","description":"https://t.co/Lhh178sIjq","followers_count":373402,"friends_count":828,"statuses_count":5674,"profile_image_url_https":"https://pbs.twimg.com/profile_images/1677873294/image_normal.jpg","screen_name":"natfriedman","location":"California","entities":{"description":{"urls":[{"display_url":"nat.org","expanded_url":"http://nat.org","indices":[0,23],"url":"https://t.co/Lhh178sIjq"}]}}},"details":{"type":"The Innovator","description":"A tech entrepreneur who funds and builds audacious, science-forward projects â from cracking 2,000âyearâold Herculaneum scrolls to testing plastics in Bay Area foods. He mixes deep technical ability with public-minded prizes and blunt, highâsignal takes on the state of tech and policy. Expect big experiments, bold bets, and lively threads.","purpose":"To use technology, incentives, and community muscle to unlock hidden knowledge and solve stubborn realâworld problems â accelerating discovery, preserving cultural heritage, and turning curiosity into measurable breakthroughs.","beliefs":"Believes in building tools not just for profit but for progress: open collaboration, meritocratic teams, rigorous engineering, and bold, prizeâdriven approaches that attract talent worldwide. Values evidence, practical impact, and transparency, and trusts engineers everywhere to deliver when given the resources and challenge.","facts":"Fun fact: Nat launched the Vesuvius Challenge and helped read Herculaneum scrolls after 2,000 years (first revealed word: 'ĎÎżĎĎĎ ĎÎąĎ'). Profile snapshot: ~279,291 followers, following 833, and ~5,674 tweets. The Vesuvius Challenge awarded a $700,000 grand prize and announced a new $100,000 prize for 2024.","strength":"Visionary projectâbuilder who can fund and mobilize top talent, translate technical complexity into public narratives, and attract media and community attention to big, otherwise-neglected problems.","weakness":"Bluntness on social media can provoke polarization; covering many disparate topics risks diluting a core audience; occasionally trades nuance for punchy takes that spark heated replies.","roast":"Youâre the only person who will bankroll a $700k archaeology prize, crow about reading ancient scrolls, and then spend the next afternoon arguing about takeout plastics on Twitter â basically Indiana Jones with a startup pitch deck and a very stubborn reply button.","win":"Spearheaded the Vesuvius Challenge that decoded parts of the Herculaneum scrolls for the first time in 2,000 years, crowning a $700k-winning team and publicly revealing never-before-seen ancient text.","recommendation":"Tell the story like a serialized documentary: post tight, image-rich threads that show stepâbyâstep progress on projects (code, scans, people), pin milestone threads, run regular Spaces/AMAs with winning teams and scholars, share short explainer videos and data visualizations, tag collaborators to amplify reach, convert interest into a newsletter or mailing list for deeper engagement, and use targeted promoted posts to turn curious viewers into longâterm followers."},"created":1774576954911,"type":"the innovator","id":"natfriedman"},{"user":{"id":"180115578","name":"Mira Murati","description":"Now building @thinkymachines. Previously CTO @OpenAI","followers_count":731576,"friends_count":620,"statuses_count":367,"profile_image_url_https":"https://pbs.twimg.com/profile_images/2057853902918455296/D_nfq15L_normal.jpg","screen_name":"miramurati","location":"San Francisco, CA","entities":{"description":{}}},"details":{"type":"The Innovator","description":"A product-minded technologist who builds infrastructure and teams to make advanced AI useful and understandable. Former CTO of OpenAI now founding Thinking Machines to push open science, strong foundations, and practical applications. Combines deep engineering chops with a people-first leadership style.","purpose":"To advance AI by creating solid technical foundations, practical tools people can actually use, and an open scientific culture that spreads knowledge â so powerful systems become widely useful, transparent, and responsibly deployed.","beliefs":"Believes in people-first engineering, open science, and rigorous foundations as the path to trustworthy, useful AI. Values collaboration, transparency, and building tools that adapt to real human needs rather than hype. Trusts that scaling capability must go hand-in-hand with clarity and responsibility.","facts":"Fun fact: Mira launched the ChatGPT iOS app and served as CTO of OpenAI before founding Thinking Machines (now @thinkymachines). She reaches a large audience (452,578 followers) and frequently tweets short, people-focused notes that highlight team and product wins.","strength":"Combines deep technical expertise with product instincts and team-building â able to translate research into shipped products, inspire engineers, and communicate complex ideas succinctly. Skilled at launching high-impact features and leading large engineering organizations.","weakness":"Can be spread thin across big ambitions; balancing foundational research, product deadlines, and public-facing commitments risks diluting focus. Public visibility invites intense scrutiny, and a people-first stance sometimes means internal trade-offs slow rapid iteration.","recommendation":"Grow on X by mixing short, high-impact product posts with explainers and behind-the-scenes threads: (1) post concise technical threads that break down one idea per thread, (2) spotlight teammates and lab wins to amplify the org, (3) share reproducible mini-demos or visualizations, (4) host periodic Spaces/AMA sessions, and (5) pin a clear intro thread about Thinking Machines' mission and how followers can engage or contribute.","roast":"You move so fast you probably have a 'launch' button for your coffee machine â and it ships with a changelog. Also, your inbox has trust issues: it assumes everything you send will become a product roadmap.","win":"Led engineering and product efforts at OpenAI that shipped major products (including the ChatGPT iOS app) and then founded Thinking Machines to scale open, foundational work â a rare blend of technical leadership, product delivery, and community-building that reshaped how people use AI."},"created":1774576830523,"type":"the innovator","id":"miramurati"}],"activities":{"nreplies":[],"nbookmarks":[],"nretweets":[],"nlikes":[],"nviews":[]},"interactions":null}},"settings":{},"session":null,"routeProps":{"/creators/:username":{}}}
