Redpoint

From the moment Deepti Yenireddy shared her vision for Boon with us, it just clicked. Satish Dharmaraj, the team at Redpoint, and I are thrilled to support this amazing group as they bring AI-powered efficiency to logistics! Hear more about Boon and why we invested from Deepti and me below!

Congrats to Deepti Yenireddy and the Boon team on their $20.5M Seed + Series A funding! Their AI workflow platform is redefining the supply chain and logistics industry—can’t wait to see them transform fleet operations and automate manual tasks off the table! https://lnkd.in/et6A-unM

🚀 Exciting news! We are thrilled to announce our $100M Series C financing! This round was led by Notable Capital, with significant participation from Redpoint Ventures, and support from Ribbit Capital, Thrive Capital and GIC. This new funding will allow us to expand our product suite, scale operations, and reach new markets. At Parafin, our mission is simple: to grow small businesses. They are the backbone of our economy, but traditional institutions often leave them behind. Through our state-of-the-art machine learning models and seamless platform integrations, we’re leveling the playing field and driving small business growth in ways never seen before. 💡 What we’ve accomplished so far: 💲 Funded nearly $1B annually for tens of thousands of small businesses across the U.S. and Canada. 📈 400% growth in volumes since our Series B in September 2022. 🚀 Launched capital programs with industry giants including Amazon, Walmart, DoorDash, and TikTok. Thank you to our investors, partners, and the incredible team at Parafin for making this possible. Together, we’re not just building financial products—we’re building a movement to grow the small business economy. Onward! 💪

Until November Bob McGrew was the Chief Research Officer at OpenAI. He’s been at the center of the exciting advances across AI and before that spent a decade leading engineering and product for Gotham at Palantir Technologies. He’s seen it all and Jacob Effron and I were fortunate to get his insights on AI and where the space is going. This was one of my favorite episodes of Unsupervised Learning to date. A few highlights: The Evolution of OpenAI 🧬 Bob shared several insider accounts of OpenAI's journey to being synonymous with the AI revolution along with various key decisions where the company had to “refound itself.” This included launching ChatGPT (“we thought if 1000 people used it then it would be successful).” He also shared really interesting insights into what makes good AI researchers. In one story, he remembered Aditya Ramesh working on Dall-E for two years and showing him a blurry blob one day (it was supposed to be a panda skating on ice) saying “it’s really beginning to work.” Aditya could see the pieces coming together and the slow but steady progress that would soon change the world and had the drive to keep pushing forward. Sora and Its Future 🎥 Bob's enthusiasm for Sora was palpable. He delved into the unique challenges of video generation, told us how the product team thought through the offering, and also what an achievement it was to democratize access to such a powerful new tool. Robotics: A Great Time to Build 🤖 Bob worked on robotics early on in his time at OpenAI and is a true believer in a coming robotics revolution, particularly in retail and work environments. He thinks it’s a great time to be starting a robotics company due to the thriving ecosystem and vision models being able to translate input to plans of action faster than ever before, and shared where he thinks the space is going. Why isn’t AI progress faster 🏗️ Bob shared a provocative thought: if in 2018 we knew what AI capabilities were today, we’d have assumed it would have changed GDP dramatically more than it has. He shared that we’ve learned there are many parts of human jobs that aren’t able to be automated and a really interesting vision for how humans and AI can best work together. Barrier to Enterprise: Scaling + Reliability 📈 Bob’s “rule of 9s” is that going from 90% to 99% reliability requires an order of magnitude increase in compute, as does 99% to 99.9% and so on, and thus 2-3 years of work for each iteration. The challenges of 2025 will be focused on scaling to larger amounts of compute, which requires innovation in unsolved problem areas across data, pretraining, hardware, and optimization. To break into enterprise settings, reliability will have to increase and a mixture of Programmatic approaches and Computer Use approaches will be leveraged to help to an extent. This was an incredible conversation - check it out below: YouTube: https://lnkd.in/gKBWewGt Spotify: https://bit.ly/401e7Kw Apple: https://bit.ly/3VJcQp0

Congratulations to Nuon on their Series A! Excited to see Nuon empower businesses through their "Bring Your Own Cloud" (BYOC) platform 🙌 https://lnkd.in/g4r3t_Ti

Fantastic segment on Bloomberg with Jason Warner, CEO of poolside!

Just one day before OpenAI’s groundbreaking full o1 release was made public, we sat down with Noam Brown, a pioneer in AI research and a key part of the o1 team. Known for his transformative AI work in planning in games like Poker and Diplomacy, Noam shared his unique perspective on how models like o1 are reshaping the future of AI—and what this means for builders, researchers, and the broader industry. Here are some of my favorite takeaways: 🔍 Are Foundation Models Hitting a Wall? His perspective is nuanced. While scaling models through pre-training has driven enormous progress, it risks becoming economically unsustainable given the costs required for another 10x improvement. Noam believes test-time compute is the next frontier and has substantial room to scale. ☕ Noam’s Coffee Chat with Ilya in 2021 One of the most compelling moments came when Noam reflected on a conversation with OpenAI’s Ilya Sutskever. At the time, Noam was skeptical about AGI in the short-term, citing a lack of general solutions for scaling test-time compute. To his surprise, Ilya agreed that this would be required in addition to scaling pre-training. Noam believed solving test-time compute would take at least 10 years—yet progress in just two or three years has shattered those expectations. 🛠 How Builders Should Use o1 For builders wondering when to use o1, Noam offered valuable guidance. While o1 excels at planning, strategizing, and tackling multi-step problems, it isn't a universal solution. He explained that builders need to match the tool to the task: o1 models shine in scenarios requiring strategic depth, while dense pre-trained models like GPT-4 remain optimal for general language tasks. His advice is to experiment, iterate, and recognize where specialized capabilities can create the most value for your specific use case. 🌐 Will There Be One Model to Rule Them All? Noam envisions a hybrid world. While models are increasingly versatile, he believes it doesn't make sense for every query to run through the full model but rather models to increasingly call tools to accomplish things like multiplying numbers. An amazing and timely conversation with Noam. Check it out: YouTube: https://lnkd.in/gfaDAjkh Spotify: https://bit.ly/4g7wC5N Apple: https://bit.ly/41m9hsx

Big announcement from poolside - their generative AI Assistant and foundation models are soon to be available as a 1st party service on Amazon Web Services (AWS). Congrats to the poolside team for another major milestone towards building next generation AI for software engineering. Full press release: https://lnkd.in/dq5DvVQT

Congratulations to Jordan Segall for being named one of Business Insider's Rising Stars of Venture Capital 2024! https://lnkd.in/gEZPn58f

Last week, Cradle announced their $73M Series B to further expand their AI-enabled protein engineering platform. So far, they’ve partnered with 21 customers and achieved cost reductions of up to 90% on R&D projects. On this week’s Vital Signs, I sat down with Cradle’s Co-Founder & CEO, Stef van Grieken. We discuss how Cradle plans to use the new funding, what’s been happening in AI x bio, the technical underpinnings of Cradle’s platform, and more. Some highlights: 🧪 The current state & future of AI x bio Stef says that we’re still in the early stages of AI x bio. He highlights three areas where AI can play an important role: 1) hit identification for easy targets via novel binders 2) structural de novo models where researchers have some understanding of the target but want to generate some variance 3) multi-property optimization to learn from experimental results and reduce the number of experimental cycles needed. Stef advocates for more experimental context in models and better benchmarks that are relevant to bio. He also talks about how the valuable datasets in bio are kept private, meaning the public datasets are inherently less valuable, which might bias models towards irrelevant directions.  🧪 How Cradle designs their models Stef shares how their foundation model has two major components: 1) A predictor component which has some knowledge of properties (e.g., stability, expression), works decently well in zero-shot, and sees all of the assay data. 2) A generator to search the local search space and is conditioned to understand the relevant domain (e.g., providing evolutionary information, providing some labeled data without leaking too much). Stef mentions that it’s easy to go out of domain in biology given the sparsity of data, so Cradle has invested significantly in model confidence around generated sequences.  🧪 What’s next for Cradle Stef explains Cradle’s three goals: 1) Most protein models to date are assuming a fixed vocabulary with natural amino acids, but Cradle wants to also represent non-natural ones. 2) Models are currently effective at optimizing a protein construct once a user has properly formatted it. Stef hopes to have models reformat from large libraries – e.g., generating a bispecific antibody instead of configuring from the individual components. 3) Cradle plans to do more zero-shot learning on their panel of assays in areas like immunogenicity. An insightful discussion on all things AI x bio! Listen to the full episode below: Spotify: https://bit.ly/3D0Cy20 Apple: https://bit.ly/3ZiSyDZ