About
My professional interests have always been "empowering people with data & technology"…
Articles by Michael
Activity
21K followers
-
Michael Li reposted thisMichael Li reposted thisfal Agent now has a $50 tier for anyone who wants to experiment. With H3 Max, a 5-second 768p video costs just $0.20 and generates in under 3 seconds. As generation gets faster and cheaper, we want to make it easier for more people to explore what’s possible. fal Agent gives you one workspace for every image, video, and 3D model on fal. It chooses the right model for each task while keeping your characters and references consistent across models. The full $50 comes back as fal credits, usable anywhere on fal, including Playground. Unspent credits roll over each month. Try it now: fal.ai/agent
-
Michael Li shared this#KDD2026 was a great one. It was my privilege to share how we have evolved our ML journey from classical ML models, to LLM powered agents, and now toward cutting edge Physical AI in line with our our vision of “Physical automation to transform industry and move the world”, at the invited talk representing ATOMS. For anyone interested, here is the full deck https://lnkd.in/gUB8eydP Besides the talk, one of the best parts of KDD is always reconnecting with old friends and colleagues, while also meeting other researchers and industry peers, exchanging ideas on the latest developments in the field, and discussing where things are headed next. And again, talent is always the key and we’re hiring! If you want to work on challenges spanning across machine learning, agentic AI, robotics, infrastructure, and real world automation, just visit our career website at https://lnkd.in/eQxxywgN and feel free to DM me with any questions! Lastly, I hope to see many of you at #KDD2027 in the heart of Silicon Valley next year! #KDD2026 #KDD2027 #ATOMS #PhysicalAI #Hiring
-
Michael Li reposted thisMichael Li reposted thisTamie Jiang was one of the hardest working, smartest, and most genuinely kind people I've had the privilege of working with at Coinbase. She left an impact on everyone who knew her, and she passed away unexpectedly, leaving her family facing costs no one should have to carry alone. If you're able to give or share, it would mean a lot to her family during an unimaginably hard time. Link below. 💙Donate to In Loving Memory of Tamie Jiang, organized by Seiji KawanabeDonate to In Loving Memory of Tamie Jiang, organized by Seiji Kawanabe
-
Michael Li reposted thisA risk score is a projection: a high-dimensional evidence space collapsed onto a single scalar. The mapping only runs one way - you cannot recover the evidence from the number, and the judgment lives in the dimensions you threw away. A regulator doesn't ask how risky it looked. A regulator asks what you can prove. A score can rank. Only evidence can judge. CipherOwl is on this bet.Michael Li reposted thisThe compliance industry is in an arms race to ship more rules, more labels, more analysts. It's the wrong arms race, and the math proves it. Stablecoins settled tens of trillions of dollars in 2025. Financial institutions burn roughly $200B a year on financial-crime compliance, most of it people tracing transactions one at a time on a ledger that never stops. Headcount grows linearly. Volume grows exponentially. TD Bank had thousands of compliance staff and still paid $3B over a 70,000-alert backlog. Nobody there was lazy. The architecture was linear, and the problem wasn't. Rich Sutton called this the Bitter Lesson: methods that scale with compute beat human-crafted cleverness, every time, eventually. Every handcrafted compliance rule is that cleverness, running on borrowed time. So here's my thesis on on-chain investigation. Part 1 of a series: - Rule-based screening is a ceiling, not a floor. - The future is agents with an on-chain world model. - Humans set the objective and render judgment. Agents run the search: every hop, every counterparty, every hour. - Output that isn't presentable, interpretable, and explainable isn't evidence. The last one matters most. A risk score is a projection: a high-dimensional evidence space collapsed onto a single scalar. The mapping only runs one way — you cannot recover the evidence from the number, and the judgment lives in the dimensions you threw away. A regulator doesn't ask how risky it looked. A regulator asks what you can prove. A score can rank. Only evidence can judge. We building CipherOwl on this bet. Part 1 - https://lnkd.in/gCRikU8h Part 2 is about why compliance is becoming computational. As usual open for discussion!The Bitter Lesson of On-Chain Truth Seeking · CipherOwlThe Bitter Lesson of On-Chain Truth Seeking · CipherOwl
-
Michael Li shared thisI'm excited to represent Atoms at this year's KDD as we begin an exciting new chapter. We’ll be giving an invited talk on Aug. 12 at 9:30 AM. As TK recently shared, our vision is "Physical automation to transform industry and move the world." (https://lnkd.in/e5XjA3vU) We believe the next frontier is automating the physical world, by bringing AI, robotics, and software together to solve some of the hardest real world problems at industrial scale. KDD has always been where many of the breakthroughs behind that future begin. I'm looking forward to meeting fellow researchers, engineers, and builders who are pushing the boundaries of machine learning and AI, and exploring how those advances can create impact beyond the digital world. If you're attending KDD this year, I'd love to connect. Most importantly, we're hiring! If you're excited about our vision and want to be part of an exceptional team building technology with tangible real world impact, please visit the Atoms careers site at https://lnkd.in/eQxxywgN. Feel free to DM me with any questions. #KDD2026 #MachineLearning #AI #Robotics #PhysicalAI #Hiring #ATOMS
-
Michael Li reposted thisMichael Li reposted thisRegistration for GenMedia Conference 2026 is now open. On September 24, 2026, we'll bring together the people shaping the future of generative media in San Francisco, from frontier AI labs and the fastest-growing startups to enterprise teams deploying AI in production. Expect a day of keynote talks, product announcements, technical deep dives, and conversations with the builders defining what's next across image, video, audio, and creative AI. GenMedia Conference is application-only to ensure a curated audience of founders, developers, researchers, creators, and enterprise decision-makers. If you're building the future of generative media, we'd love to see you there. Register your interest today for a chance to attend.
-
Michael Li reposted thisMichael Li reposted thisWe are proud to introduce Ego1, our first egocentric capture headset for Physical AI. We co-designed the hardware and the perception stack to turn ordinary first-person activity, especially manipulation, into training data for robotics foundation models. It is all-in-one: two synchronized cameras and a 200 Hz IMU, recording on its own with no phone or laptop. Those raw streams become the outputs robotic models train on: a head trajectory, 21-keypoint 3D hand tracking for both hands, and dense depth. See the write up:Introducing Ego1: an all-in-one headset for egocentric robot-learning dataIntroducing Ego1: an all-in-one headset for egocentric robot-learning data
-
Michael Li shared thisAmazingly insightful report from fal on Generative Media! Strongly recommended!Michael Li shared thisToday we’re launching the State of Generative Media report: a comprehensive look at how the GenMedia space evolved this year and where it’s heading next. In this report, we break down: • The models that defined 2025 across image, video, audio, and 3D • Enterprise adoption patterns and real-world use cases • Developer usage trends across modalities • Where the ecosystem is heading in 2026 What changed this year wasn’t just quality. It was scale; continuous model releases, real-time performance gains, and expanding production workflows across industries. Read the full report here: https://lnkd.in/gtUW5AnR
-
Michael Li reposted thisMichael Li reposted thisWe’re thrilled to welcome Philippe Clavel as Chief Technology Officer at Kikoff! https://lnkd.in/gGEb6u_Y A seasoned technology leader and founder, Philippe brings deep experience building and scaling consumer platforms, leading high-performing engineering teams, and developing AI-powered solutions that make a real-world impact. His people-first approach and passion for innovation will be instrumental as Kikoff continues to strengthen its AI leadership and expand access to affordable credit. As we enter our next chapter of growth, Philippe will help drive our mission to build technology that simplifies finances, lowers costs, and empowers consumers with smarter, fairer financial tools. Welcome to the team, Philippe. We’re excited for what’s ahead!
Michael Li commented on a post
1mo
Great piece! Computation will be everywhere with AI, compliance won’t be an exception.
-
Michael Li liked thisMichael Li liked thisAnother deep dive on the engineering side, this time from Javier on how our new ingestion and query protocol - QWP - sends timestamps over the wire. Rather than the full 64-bit value, QWP uses Gorilla encoding from Facebook's 2015 TSDB paper and sends how much the gap between rows changed. On a steady feed that's zero, which costs one bit, so a regular series gets ~57x smaller while bursty crypto trades land at ~2.5x. If a value doesn't fit, the encoder falls back to plain 64-bit for that column in that batch, so it never makes anything bigger. Blog post in the first comment. #questdb
-
Michael Li liked thisMichael Li liked thisOur researchers just released Physis-Lang, an open self-evolving framework that adds physics reasoning to video captions. The captions explain why and how a scene unfolds. We use them to fine-tune world models and add that reasoning to prompts when generating video. Adding physics reasoning to the prompt alone improved NVIDIA Cosmos 3's PhyGenBench score by 5.62 points without retraining. Physis-Lang with NVIDIA Cosmos 3 Super and Nano now rank #1 and #2 on the Physics-IQ Verified image-to-video leaderboard! Paper and project: https://bit.ly/4AGFLgf #worldmodel #physicalai #nvidia
-
Michael Li liked thisMichael Li liked thisExcited to announce Living Science! Papers have been the de facto unit of knowledge in science. But science is not static, and a paper shouldn't be the final destination—especially in empirical research. With AI agents, we're revisiting influential research as new data, models, and methods arrive, starting with economics. We've published three extensions of economics papers we admire. For example, our extension of Cengiz et al. (2019) examines 236 state-level minimum wage increases through 2024, up from 138 in the original study. Across those increases, affected workers' wages rose 4.6%. The estimated employment change was −1.1%, statistically indistinguishable from zero. We use SAI Labs's replication agent to try to reproduce key findings, then extend the analysis to newer data where possible. We sanity-check the results and ask for feedback from the original authors. Each extension includes code and data, a chat interface to help readers understand and interrogate the results, and an invitation for public feedback. Why build a shared platform when AI increasingly lets researchers revisit papers on their own? • To help people build on existing analyses instead of starting from scratch. • To make new evidence common knowledge. An update on one person's computer does little to change what everyone else keeps citing. • To support science as a communal effort, with a shared public record where people can catch errors, debate interpretations, and contribute updates. This is a joint effort with Evan Rose, made possible by Kenneth C. Griffin Applied Economics Incubator and SAI Labs. We plan to add at least one paper a week. Help us keep research alive: check our analyses, join the discussion, or nominate a paper to revisit. https://livingscience.ai/
-
Michael Li liked thisMichael Li liked thisToday we are excited to announce HeyGen Video, HeyGen's first general-purpose video model, built for businesses that need production-quality video without production-level costs. One model covers text to video (T2V), image to video (I2V) and reference to video (Ref2V): start from a prompt, a first frame, or your own reference images. Every clip comes back with sound (dialogue, ambience, and effects) generated in the same call. In blind side-by-side tests, it performs alongside the top video models at a fraction of their price per second. That makes it practical for everyday business work: product demos, training, onboarding, and property walkthroughs. Pricing starts at $0.01 per second through October (50% off the standard $0.02). It is live now on the HeyGen API and through OpenRouter, Runware, and ComfyUI. Built on MiniMax H3 and post-trained by HeyGen. Thanks to MiniMax for the partnership. Try it: https://lnkd.in/g_Mk_7NC See the benchmark and 100+ clips: https://lnkd.in/gvVVQ3NT
-
Michael Li liked thisMichael Li liked thisI didn’t expect to get into robotics. I did it because I saw how robotics could change physical labor industries forever. I have been in Computer science for more than 15 years. CS had always interested me because there is always so much left to build. Two or three years ago, I still doubted whether AI would advance far enough to make general-purpose robotics possible. By the end of 2025, progress in AI had changed my view. Robotics is downstream of AI, and general models for robots started to look achievable. And one of the tipping point is I saw Physical Intelligence's demo at Neurips 2025. Long horizon task, fully autonoumous. I went to a few robotics events and met people in the industry. The field was open enough that a new company could still help define how things worked. I felt I could compete. Since then, robotics has moved even faster than I expected. I think generalist foundational models will arrive fairly soon. This will eventually lead to a future where many more robots working alongside people in factories and other physical workplaces. The best way companies built around physical labor can prepare for that future is by owning their data infrastructure. Physical AI needs more capable foundation models, and those models need far more high-quality data. Data is the bottleneck today. There are six layers in Physical AI: 1. Data collection hardware 2. Data operations 3. Data cleaning and labeling 4. Pre-training 5. Post-training 6. Deployment The compnay, General Intelligence Labs, I cofounded with Haichen Shen and Junyuan Xie focuses on the first three layers so that companies can own the data infrastructure behind everything that follows. In the short term, a company with that infrastructure can collect data and sell it to model builders. It becomes an additional revenue stream. In the long term, that company can own its data flywheel and use it to build soverign physical AI for its own business. A company that waits will watch its competitors adopt robotics first, then have to catch up. We aim to be the primary data partner for companies building intelligent robot models. We also want to help companies in physical industries build physical AI into their own businesses. It will probably take 5 to 10 years for this transition to fully go mainstream, but the early adopters are moving right now! cc Saharsh Chordia Robotics will move faster than most people can perceive. Companies built around physical labor have less time to prepare than their leaders expect.
-
Michael Li liked thisMichael Li liked thisAnother security incident last weekend... ~$380M+ was stolen from Bitget Bitget asked THORChain (a decentralized network) to refuse service to the attacker's addresses. THORChain declined. 34 hours later, the stolen ETH started swapping into Bitcoin through it. In about 45 hours, the attacker pushed 40,551 ETH through THORChain in 1,485 deposits and came out with 1,238 BTC, roughly $103M, spread across 36 Bitcoin addresses. We traced every one of those deposits. A few things stood out: • $60.5M of USDT, USDC, and tokenized gold was sold for ETH in 21 minutes. Stablecoins can be frozen. ETH can't. • 68,923 ETH sat untouched for days, then was peeled through 377 wallets in about 56 hours. • One Bitcoin address collected the proceeds of both stolen XRP and stolen ETH, which ties the two legs of the theft together on-chain. What compliance teams should take from it: every THORChain deposit carries its Bitcoin destination in plain text, before any BTC is paid out. And every Bitcoin payout carries the hash of the Ethereum deposit that funded it. That makes the ETH-to-Bitcoin link provable from public data alone. One catch. 39% of the ETH never touched THORChain's router. It went straight to vault addresses with the memo in the transaction data, so a monitor that only watches the router would have missed it. Strix, our investigation agent, ran this trace in CipherOwl Workbench as the funds moved. The session is public, and the write-up lists the addresses so exchanges can screen against them. Full breakdown and the live session are in the comments. #AML #CryptoCompliance #BlockchainAnalytics https://lnkd.in/g9yU2xsyBitget hack: how 40,551 ETH moved through THORChain to Bitcoin · CipherOwlBitget hack: how 40,551 ETH moved through THORChain to Bitcoin · CipherOwl
-
Michael Li liked thisMichael Li liked this
-
Michael Li liked thisMichael Li liked thisExcited to see NVIDIA and Sea deepen our partnership, with Sea becoming the first enterprise in Southeast Asia to adopt the NVIDIA Vera Rubin platform. https://lnkd.in/gf9MqSNB Christopher LowVerdi MarchVikram RaoAzita MartinYogesh Agrawal
-
Michael Li liked thisMichael Li liked thisIt was great sharing the stage with leaders from Impact AI, Blackrock, and Visa to discuss The AI Proving Ground at WeAreDevelopers last week. Among other shared sentiments, its clear that rigorous evaluation of AI pilots prior to deployment is a common standard.
Experience & Education
Publications
-
Ranking the 100 Most InDemand Employers Using LinkedIn Data
blog.linkedin.com
See publicationAt LinkedIn, data is in our DNA, and we love using it to help our members and companies better understand the professional landscape. Take a look and see if your company made one of our lists at talent.linkedin.com/indemand.
-
Potential impacts of CF3I on ozone as a replacement for CF3Br in aircraft applications
Atmospheric Chemistry and Physics
Projects
-
Merlin
Leverage LinkedIn's unique big data assets to develop advanced analytics tools for our sales and marketing organization.
Other creatorsSee project
Languages
-
English
Native or bilingual proficiency
-
Chinese
Native or bilingual proficiency
Recommendations received
2 people have recommended Michael
Join now to viewOther similar profiles
Explore more posts
-
AJ Ghergich
CMOs are all asking me the same question: Do we have the right AI search strategy to actually see success in 2026? That’s exactly what I spoke to María José Gutiérrez Chávez about for Inc. Magazine on how AI is changing last‑minute gift buying. This is how I see it: 2025 was the year of AI visibility and getting everybody on board. 2026 will be the year of AI action. The new rules for brands: 🤖 AI is the gatekeeper for last-minute purchases 🔍 If AI can't read it, AI can't recommend it ⏱️ High-intent moments now collapse into 3-4 AI-picked options 🏆 Winners treat AI as a primary customer, not a side channel Check out the full piece here: https://lnkd.in/gt6eeJ9h #SEO #AI
7 Comments -
Ryan Vanshur
Endgame just dropped their 2026 AI Benchmark Report and it's one of the best GTM data pieces I've read in a while. 31,000+ real AI interactions. Not survey data. Not "how do you feel about AI." Actual behavioral data from hundreds of GTM teams. The whole report is worth your time, but here's the stat that jumped out at me: 💡 22.8% of all AI activity is Account Intelligence Most people will read that and think "sales research tool." I think it's the unlock that makes everything else possible. Here's what I mean...AI adoption in GTM doesn't happen all at once. It moves in waves. And Account Intelligence is the forcing function that breaks the first wave open. Nobody wakes up and says "let's go all-in on AI across the revenue org." They have a pain point. Usually it's pre-meeting research or account context. They solve that one thing. The skepticism drops. The hallucination anxiety fades. And then something interesting happens. The jobs-to-be-done start migrating downstream. We're living this at Handle.com right now. We started with account intelligence. Once that was working, the prompting behavior drifted toward deal advancement. Then pipeline monitoring. Now it's pushing into post-sale and renewal workflows. The wave is real. At least for us. Two other things from the report that deserve more attention: #RevOps is the most leveraged persona in AI. 14% of users driving 53% of Pipeline Inspection activity. That's 4x their expected share. If you're a RevOps leader and you're not looking at this data, you should be. Teams spend 2x more AI time understanding accounts than producing deliverables. That's not a productivity story. That's a context story. The highest-value use of AI isn't "write me an email." It's "help me understand what I'm walking into." 👇 Full report linked below. Worth 15 minutes of your day. https://lnkd.in/g-BzFifN
1 Comment -
Nagesh Singh Chauhan
Why MCP Matters for the Future of AI As LLMs move from demos to real-world systems, the biggest challenge isn’t model capability—it’s context, action, and integration at scale. In this blog, I break down Model Context Protocol (MCP): • Why prompt-based integrations don’t scale • How MCP solves the M×N integration problem • Why MCP is emerging as the operating system for AI agents If you’re building agentic systems, RAG pipelines, or enterprise AI, MCP is a concept worth understanding early. 🔗 Read the blog here: https://lnkd.in/gJpcAHdt #AI #LLM #MCP #AgenticAI #Architecture #FutureOfAI
-
Andrei Lopatenko
An interesting approach to the multi-turn chart editing task. LLM-supported editing of complex artifacts will be one of the major applications of LLMs, embedded within broader consumer and business workflows. As a result, evaluation methods for these types of interactive, multi-step editing tasks will become increasingly important and widely adopted. https://lnkd.in/gThNvenR
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content