About
Early stage tech venture capitalist @Mosaic Ventures
Previously:
- Investing:…
Previously:
- Investing:…
Articles by Simon
Activity
5K followers
-
Simon Levene shared thisvery excited about this launch and that Nexar Inc.'s time has comeSimon Levene shared thisWhen I agreed to join as the CEO of Nexar Inc., it was because I believed that Nexar was uniquely positioned to become the intelligence layer for all Physical AI. That meant grounding machine intelligence in the real world, so systems operating in it can be measured against truth, not simulation. Roads are one of the clearest places that matters. 1.35 million people die on them every year, and one reason that number has been so hard to move is that vehicles still struggle to understand what is about to happen before it is too late. Today, I could not be more excited to launch BADAS 2.0, the world’s best collision prediction model, outperforming frontier benchmarks like NVIDIA Cosmos and Google Gemini. BADAS 2.0 is built on a V-JEPA2 world model architecture and trained entirely on real-world video from our network capturing over 100M miles every month. Long-tail edge cases. Rare events. Just real-world chaos. The result is a model that doesn’t just detect events. It understands what is happening, explains it, and generalizes to environments it has never seen before. We published the benchmarks. We shared the methodology. Check it out! You can even try it on your own videos: https://bit.ly/4emDi1M Let's go!! #CollisionPrevention #PhysicalAI #RoadSafety #AutonomousVehicles #FleetSafety
-
Simon Levene reposted thisSimon Levene reposted thisOur firm is led by a small partnership – and powered by data infrastructure and a set of AI-native workflows. Mikhail, our Head of Data and Product, shares what's under the hood: https://lnkd.in/gSnHVYxu
-
Simon Levene shared thisvery excited for the next chapter Arto Yeritsyan and team AsyncSimon Levene shared this📢 𝗣𝗼𝗱𝗰𝗮𝘀𝘁𝗹𝗲 𝗶𝘀 𝗻𝗼𝘄 𝗔𝘀𝘆𝗻𝗰. This isn’t a change of direction. It’s just the clearest name for the mission we started with on day one. We began as a frontier 𝗔𝗜 𝗹𝗮𝗯 with a simple goal: give voice to everyone. Over the past few years, we’ve trained our own models and built the tools that millions of creators now use to ship real content. But the next inevitability is already here. 𝗩𝗶𝗱𝗲𝗼, 𝗮𝘂𝗱𝗶𝗼, 𝗮𝗻𝗱 𝘃𝗼𝗶𝗰𝗲 𝗮𝗿𝗲 𝗰𝗼𝗻𝘃𝗲𝗿𝗴𝗶𝗻𝗴 - and agents are going to automate the messy middle. Voiceovers, dubbing, subtitles, localization, generations... today it’s all still too fragmented across different tools. 𝗔𝘀𝘆𝗻𝗰 𝗶𝘀 𝗼𝘂𝗿 𝗮𝗻𝘀𝘄𝗲𝗿. ⚡️ It’s the 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄 𝗢𝗦 for Video + Audio + Voice - from intent to production. No more separate features or generation in silos. Just one project, one timeline, and one prompt that handles everything: 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲 -> 𝘁𝘄𝗲𝗮𝗸 𝘄𝗶𝘁𝗵 𝗔𝗜-> 𝗮𝗱𝗱 𝘃𝗼𝗶𝗰𝗲𝗼𝘃𝗲𝗿-> 𝗱𝘂𝗯 -> 𝘀𝘂𝗯𝘁𝗶𝘁𝗹𝗲𝘀 -> 𝗰𝗹𝗶𝗽𝘀 -> 𝗯𝗿𝗮𝗻𝗱 𝘃𝗼𝗶𝗰𝗲. In this next era, you don’t just generate files - 𝘆𝗼𝘂 𝘀𝗵𝗶𝗽 𝗼𝘂𝘁𝗰𝗼𝗺𝗲𝘀. Create once. Localize everywhere. Repurpose continuously. One thing is certain: Voice is no longer only for humans. It’s becoming the 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 𝗳𝗼𝗿 𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲. 🧠 We're moving forward with two clear paths: 🚀 𝗔𝘀𝘆𝗻𝗰 𝗖𝗿𝗲𝗮𝘁𝗼𝗿: https://lnkd.in/d7jgD3DE 𝗔𝗻 𝗔𝗜-𝗻𝗮𝘁𝗶𝘃𝗲 𝘄𝗼𝗿𝗸𝘀𝗽𝗮𝗰𝗲 𝗳𝗼𝗿 𝗩𝗶𝗱𝗲𝗼, 𝗔𝘂𝗱𝗶𝗼, 𝗮𝗻𝗱 𝗩𝗼𝗶𝗰𝗲. A chat-based workflow where you direct the AI to handle the entire production on a single timeline. 💻 𝗔𝘀𝘆𝗻𝗰 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿: https://lnkd.in/dqigQYNF For engineers building voice-first products - real-time streaming TTS (low latency) with more capabilities coming soon. 𝗜𝗳 𝘆𝗼𝘂’𝗿𝗲 𝗮 𝗣𝗼𝗱𝗰𝗮𝘀𝘁𝗹𝗲 𝘂𝘀𝗲𝗿: 𝗻𝗼𝘁𝗵𝗶𝗻𝗴 𝗯𝗿𝗲𝗮𝗸𝘀. Same account, same team, same product - just accelerating under a name that finally matches where we’re going. To everyone who supported us on the Podcastle journey - thank you. To our team - you built this moment. More coming soon. - Arto #GenerativeAI #AItools #CreatorEconomy #TextToSpeech #AsyncAI
-
Simon Levene shared thisSimon Levene shared thisWe tried to break OpenAI’s models. Here is what happened. We pushed leading models into long and complex conversations to test a simple question: Does AI really lose its memory under pressure? The answer was clear. The large models stayed stable, even when we overloaded the context. These are the questions that drive us at Parloa, as we want to deliver amazing AI <> Human experiences. This is why we are launching Parloa Labs! Every day our agents run millions of interactions. That gives us something unique, the ability to see how AI actually behaves in production. In our second recent study, we leveraged just that to study not only how models speak but how they decide. Our insight: We could see that GPT 5.2 reduced routing failures by 4x and executed actions much more reliably. The results of both these studies are interesting on their own. But what excites me even more is the broader mission: Open the Black Box of AI to everyone. In our view this is how we make it both trustworthy and perhaps also relatable. Thank you team for pushing the boundaries: Akshay Kumar, Anjana Vasan, Ciaran O'Reilly Ibañez, Galuh S., Mariano Kamp, Masashi Beheim, Matthäus Deutsch, Nuno Filipe Vieira Marques, Pedro Castillo, Rouven Glauert, Stefan Ostwald, Tomas G. This is literally Day One for Parloa Labs. Back to building AND discovering! Link to our research papers in the comments.
-
Simon Levene posted thisgreat conversation between Kaarel Kotkas and Gustaf Alströmer 🇺🇦 covering a lot of ground on Veriff's progress. Gustav led the investment for YC before Mosaic Ventures led the Series A https://lnkd.in/eR_gcSK4
-
Simon Levene shared thisour portfolio company async (aka podcastle) is entering the *hot* TTS space. worth checking outSimon Levene shared this📣 Following their acquisition by Meta, Play AI has closed its API and will be ending all services by 12/31. 📣 Great news for the PlayAI team, but disappointing if you’re a developer looking for a quality TTS API. But that’s why we built Async… If you’re interested in testing the quality/latency ratio of our text-to-speech API, we will happily give you free credits to get started (and btw, we already have the most developer-friendly pricing on the market 🤫 ) Just DM me and our dedicated support can help you with migration and on-boarding. Link in the comment 👇
-
Simon Levene shared thiscongrats martin, mihkel and all at Lightyear on the Series B raise https://lnkd.in/eQTgH3qgLightyear Raises $23M Series B, CEO Martin Sokk Says European Investment Needs to Be FixedLightyear Raises $23M Series B, CEO Martin Sokk Says European Investment Needs to Be Fixed
-
Simon Levene shared thiswell worth reading from benedict, as alwaysSimon Levene shared thisEvery year, I produce a big presentation exploring macro and strategic trends in the tech industry. New in May 2025, ‘AI eats the world’. https://lnkd.in/gRQM2rm
-
Simon Levene shared thisvery excited about the async launch. congrats to Arto Yeritsyan and teamAsync Voice AI: High-quality text-to-speech, designed for developers | Product HuntAsync Voice AI: High-quality text-to-speech, designed for developers | Product Hunt
Simon Levene commented on a post
1mo
welcome Rob
Simon Levene commented on a post
2mo
we havent spoken in many years but congrats Gavin, what a great outcome after 25 years! excited to see what comes next :)
Simon Levene commented on a post
2mo
great post Eran Shir Zach Greenberger excited for the year ahead
-
Simon Levene liked thisSimon Levene liked thisImportant day for Auterion and Airlogix. Germany and Ukraine execute on a partnership for production of Ukrainian drones on German soil at industrial scale, thanks to Chancellor of Germany Friedrich Merz and Ukraine President Volodymyr Zelenskyy. Auterion is bringing the guidance software, Airlogix the drone. Read the full news, link in comments.
-
Simon Levene liked thisSimon Levene liked thisVLMs predict the next token. World models predict what happens next in the world. That distinction is everything. BADAS 2.0 is built on V-JEPA2, the architecture Yann LeCun has argued is the right foundation for physical AI. Instead of reconstructing pixels, it learns latent-space representations of future states. It doesn't pattern-match on visual similarity. It models causality. Trained on Nexar's corpus, 100M+ real-world miles captured every month by 350K cameras, the model didn't just learn driving, it learnt physics. BADAS 2.0 generalises beyond anything we specifically trained it on: drones, forklifts, robot vacuums, construction sites, dinosaurs, X-Wing Fighters, GTA ... 99.4% AP, 4.6% false alarm rate, 91% early warning recall. And even our smaller 22 million parameter BADAS 2.0 Flash Lite outperforms NVIDIA's COSMOS fine-tuned on the exact same 2 million driving clips, which operating at 2 billion model parameters. That's not a size optimisation story. It's what happens when the right architecture meets the right data at scale. Benchmarks, methodology, and per-category breakdown are published. Run it on your own videos (including those X-Wing Fighters!). Link in the first comment. #PhysicalAI #WorldModels #CollisionPrevention #BADAS
-
Simon Levene liked thisSimon Levene liked thisToday, I’m thrilled to announce that we are significantly expanding our Journal Home partnership with BMJ ! Back in my days as an MD scientist, BMJ was always one of the gold standards. It was the place where you went for the research that truly changed clinical practice. So, seeing our partnership with BMJ grow from a small pilot into a major, long-term collaboration is a very special milestone for me personally. After a successful initial couple of years, we are now bringing the majority of BMJ’s prestigious portfolio – over 60 journals – directly to the 25 million+ researchers on ResearchGate, including many of the most influential titles in medicine. BMJ has been a leader in medical publishing for over 180 years, and it is a real privilege to help them continue to be at the forefront of engaging the global medical community. By activating these journals in Journal Home, they’re making it easier than ever for the medical community to stay informed and connected. A massive thank you to the team at BMJ for their vision and partnership, and to the ResearchGate team for making this happen!
-
Simon Levene liked thisSimon Levene liked thisWe just watched someone edit a full video in 5 minutes. It used to take them 7 days. This isn't a workflow hack or a faster export button. This is what happens when you replace the entire editing interface with a conversation. You describe what you want. Async builds it. You iterate in chat. No timeline scrubbing. No menu diving. No switching tools. This is what we've been building at Async - the Lovable for video. A single chat where you go from idea to finished video without ever leaving the page. The AI video space has been obsessed with generation. Who can render the most realistic clip. Which model wins on benchmarks. We took a different bet: the real problem was never the model. It was the workflow. 7 days → 5 minutes. That's the bet paying off. Full walkthrough in the comments 👇
-
Simon Levene liked thisCongrats Luc and team!Simon Levene liked thisThere is a gap between how AI systems are trained and where they are expected to operate. Most are trained on simplified or synthetic environments. Then deployed in the real world. That gap is where systems fail. BADAS 2.0 closes it. A V-JEPA2 world model trained entirely on real-world video, designed to learn how the world behaves, not just how it looks. This changes two things. First: generalization. The model is not limited to the distribution of its training data. It carries a representation of how events unfold, allowing it to handle scenarios it has not explicitly seen. Second: deployment. Because the model learns structure instead of memorizing patterns, it does not rely on scale to perform. That makes it possible to build lite versions that run from cloud to edge, including on CPU. We published the benchmarks and the methodology. This is what it looks like when models are trained on reality instead of approximation. See how BADAS 2.0 performs on your own videos, run them here: https://bit.ly/4cdU9T7 #CollisionPrevention #PhysicalAI #RoadSafety #AutonomousVehicles
-
Simon Levene liked thisSuper exciting milestone for Vortexa 🚀 This financing strengthens our foundation as we scale the intelligence layer for global energy and freight markets. Huge thanks to the Oxford Finance LLC team - especially Austin Szafranski and Christopher Herr - our investors, particularly the Morgan Stanley Expansion Capital team (Lincoln Isetta, Jon Sheets, Akhil Madan and Pete Chung), and everyone at Vortexa, especially Paul Harris, who made this happen. Most importantly, thank you to Vortexa's clients and partners around the world - your trust and collaboration make all of this possible 🙏Simon Levene liked this🚀 We are pleased to announce Vortexa has closed a $50M strategic debt financing with Oxford Finance LLC. This complements our Series C financing led by Morgan Stanley and further strengthens the company’s capital base for its next phase of growth. As global energy and freight markets become increasingly complex and volatile amid rising geopolitical tensions, the need for real-time intelligence in a fast-moving world has never been greater. Vortexa enables market participants to navigate shifting trade flows and freight dynamics with greater speed, clarity and confidence. Read the press release: https://ow.ly/IrN150YIOqS
-
Simon Levene liked thisSimon Levene liked thisOur next step https://lnkd.in/e6QpYFxjArtnet and Artsy Come Together Under Shared Leadership | Artnet NewsArtnet and Artsy Come Together Under Shared Leadership | Artnet News
-
Simon Levene liked thisProud to be part of this BADAS team!Simon Levene liked this“Break!!! Collision detected”. That’s the output of most AV collision prediction models, only they provide it far too late. No more. We just built the best collision prediction model on the planet. Period. BADAS 2.0 is live. 99.4% average precision. Trained on 2 million real-world driving clips. Zero synthetic frames. We're publishing our benchmarks, sharing our data, and inviting direct comparison. Built on V-JEPA2, the attention-based architecture behind Yann LeCun's vision for physical AI, BADAS 2.0 is a true world model. It doesn't match patterns. It has internalized physics, causality, and consequence. It understands reality, not just the road. It explains its predictions through attention heatmaps. It describes the unfolding situation in natural language. It generalizes across the full long tail of risk: fog, snow, animals, infrastructure, even out-of-distribution situations that have nothing to do with driving. And it runs from cloud to edge, including CPU-based devices. Three model sizes: 300M, 86M, and 22M parameters. Our smallest model outperforms NVIDIA's 2-billion-parameter Cosmos-Reason2 while being 91× smaller. What does this mean? A world where the car understands what's about to go wrong, and knows how to avoid it. The benchmarks are public. The bar is set. Try your own videos on BADAS 2.0, run them here: https://bit.ly/CMBD2I #PhysicalAI #CollisionPrediction #OpenChallenge #ADAS #VJEPA2 #Nexar
Experience & Education
-
Async
View Simon’s full experience
See their title, tenure and more.
or
Languages
-
French
-
Recommendations received
1 person has recommended Simon
Join now to viewExplore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore MoreOthers named Simon Levene
16 others named Simon Levene are on LinkedIn
See others named Simon Levene