AI21 Labs

Andrej Karpathy recently warned that we're getting "way too excited" about autonomous AI agents. As he puts it: “If I’m just vibe coding, AI is great, but if I’m trying to really get work done, it’s not so great to have overreactive agents.” This is a prime example of the challenges that arise when AI moves from casual use to serious work. So, what's the solution? In Karpathy's words, we need to "put AI on a leash." It's time to rethink how we design and architect AI systems. AI must guarantee that explicit user requirements are fulfilled, break down work into manageable steps, use the right tools for each step, and validate the output of every step. We believe that Karpathy's leash isn't a shackle, but rather a way by which enterprises can start to trust AI. What do you think? Let us know in the comments below. 👇

🚀 Jamba is now live on Kaggle! We're excited to bring the latest version of AI21 Labs' Jamba family of open models, built with a novel hybrid Transformer-Mamba architecture, to Kaggle's 25M+ community of ML developers, researchers, and learners. 🧠 This new version builds on the SSM-Transformer hybrid foundation, 256K context window and efficiency gains of previous releases, while introducing upgrades in instruction following and grounding. You can now try Jamba 1.7 directly in a Kaggle notebook, no setup required. 🎯 Whether you're benchmarking, prototyping, or just experimenting, we can’t wait to see what you build. Tag us or share your projects! 👉 Try it on Kaggle: https://lnkd.in/ejKmetX5 📦 Also available on Hugging Face and our AI21 SaaS platform. #AI21 #Jamba #Kaggle #LLMs #MoE #TransformerMamba

🚨 Agent Hijacking Alert! 🚨 In this Yet Another AI Podcast (YAAP) episode, Zenity CTO, Michael Bargury, breaks down how AI agents get hijacked and what you can do to prevent yours from going rogue. Yuval Belfer sits down with Michael to explore: 🤖 How easy it is for popular agent systems like Copilot and Cursor to be hijacked ⚠️ The sneaky ways zero-click attacks can infiltrate your AI ⛔ Why "ignore previous instructions" still poses a serious risk 💥 How one rogue agent can infect others 🔓 Real world examples of attacks: from social media triggers to CRM leaks and logic bombs 📊 The must know basics of AI observability: logs, reasoning traces, and more The new rule? Build like your AI will go rogue, because one day, it just might. 🎧 Listen now on Spotify, Apple and Youtube at the links in comments. #AI #CyberSecurity #AIrisks #AIagents #ZeroClickAttacks #AItools #TechPodcast #AIsecurity

Founders building with AI, this one's for you👇 In a recent episode of AI Basics, Jason Calacanis (This Week in Startups) in partnership with Google Cloud sat down with Yoav Shoham, co-founder and co-CEO of AI21 Labs, to explore some of the most pressing challenges in AI today. They discussed: • Why reliability remains a major hurdle for enterprise AI • What orchestration truly means and why LLMs alone aren’t enough • The risks of “agent-washing” in AI development • Small vs large models, agent to agent protocols and where real opportunities lie If you're navigating hallucinations, chasing automation, or exploring multi-agent workflows, this episode is a must. Check it out and let us know what you think! 📽️ https://lnkd.in/e8_ktDxd

✨ Maestro Now Available to All Enterprise Customers! ✨ This week at RAISE Summit, our co-founder and co-CEO Ori Goshen, announced that Maestro is now open to all enterprise customers. Previously available only via a waitlist, businesses of all sizes can now take advantage of Maestro to streamline AI solution development, enhance accuracy and drive real enterprise impact. At AI21 Labs, we believe that even the most complex AI systems should be effortless to build, and above all, highly valuable in real world business operations. That’s why Maestro was designed with the principle of 'Complex AI, Simple Assembly', to make AI not just powerful, but truly accessible, dependable and easy to deploy. We’re excited to see what enterprises will build with Maestro AI agents! #RaiseSummit #AI21Labs #AI21Maestro #EnterpriseAI #ComplexAISimpleAssembly Henri Delahaye | Michael Amar 🎗️| Hadrien de Cournon

Yesterday at the RAISE Summit, Asaff Zamir, VP of Business at AI21 Labs and Yuval Belfer sat down for an insightful fireside chat. They explored what it really takes to build and deploy custom AI systems that deliver measurable value for organizations. Key Takeaways 🔹 The “ChatGPT syndrome” and how many organizations expect workplace transformation simply by plugging in an LLM 🔹 The underground AI movement of employees using GenAI tools without org-wide direction 🔹 Why success depends more on people and workflows than on models or tech alone The promise of enterprise AI is much bigger than just efficiency gains. Enterprises can and should harness AI to tackle bigger challenges, automate and achieve more ambitious goals. The shift to custom AI systems is a critical component that will empower businesses to create solutions that drive measurable value and real-world impact and we're excited to be partner on this journey with many of you!

🔨 Complex AI, Simple Assembly. We believe that even the most complex AI systems, like great designs, should be effortless to assemble. That’s why Maestro AI agents are designed to be both simple to build and highly valuable in real world business operations. We’re on a mission to make AI not just powerful, but truly accessible and dependable. Previously available only through a waitlist, Maestro is now open to all enterprise customers. This means that businesses of all sizes can now harness the power of AI21’s Maestro to streamline AI solution development, increase accuracy and drive real enterprise impact with ease. 👩💻 Because building AI should be simple. Too many AI projects stall because organizations are forced to choose between two bad options: hoping LLMs get things right (prompt-and-pray) or hard-coding every step into brittle, rigid workflows. Neither scales. Maestro changes this by automating the complexities of AI implementation, enabling organizations to deploy truly reliable AI solutions faster and more efficiently than ever before. From financial research to contract analysis and RFP responses, Maestro agents handle it all with proven accuracy gains, particularly in complex multi-step, multi-requirement tasks. Click through our easy to follow guide to see how it works. 🔨 Don’t build alone. Let us help accelerate your time-to-value with AI21 Labs’ delivery & assembly services. Reach out to learn more! #AI21Labs #EnterpriseAI #AI21Maestro #ComplexAISimpleAssembly RAISE Summit

🎉 Live today: Jamba 1.7, the latest update to our open model family! Jamba 1.7 builds on its predecessor's novel SSM-Transformer hybrid architecture, 256K context window, and efficiency gains, while introducing improvements in grounding and instruction-following.  🔹 Why grounding? 🔹 As builders of AI for the enterprise, we know how critical factuality and contextual faithfulness are in business applications. Grounding allows models to understand and utilize a given context effectively, making them invaluable for tasks like question answering, retrieval-augmented generation (RAG), and internal search assistants. 🔹 Why instruction-following? 🔹 Like grounding, the ability for enterprise developers to control how a model operates is invaluable, from achieving accurate output to ensuring the safety of the system. Available on AI21 Studio or downloadable from Hugging Face 🤗 More on what we improved & why in the link in comments.

Yoav Shoham is a professor emeritus at Stanford University, the author of the The AI Journal Classic Paper Award for his 1993 paper on agent-oriented programming, and coauthor of Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations, a standard textbook in the field. He also happens to be our co-founder and co-CEO. With such a rich background, it’s no surprise that he’s deeply involved in the conversation about AI agents. In his latest article for MIT Technology Review, Yoav explores the growing hype around AI agents, addresses the risks of "agent-washing," the lack of clear definitions and the need for realistic expectations. While the potential for AI agents is huge, Yoav emphasizes that thoughtful design, reliability and real world application are key. At AI21 Labs, we’re already making strides with solutions like Maestro, focused on enterprise reliability. Read the full article to see why Yoav believes the future of AI agents hinges on clear definitions, structured systems and a healthy dose of realism. https://lnkd.in/d4BE2qec

✈️ The AI21 team is heading to Paris for RAISE Summit next week! We’re excited to showcase how enterprises can drive real world impact with secure, scalable AI. Stop by Booth #19 to learn how we’re helping businesses transition from AI pilots to production at scale. Be sure not to miss these key sessions: 🔹 Our co-founder and co-CEO Ori Goshen on the main stage, July 8th at 2:20 PM 🔹 A fireside chat with Asaff Zamir, VP of Business at AI21 Labs, moderated by Yuval Belfer, on July 9th at 2:00pm They’ll discuss how to build custom AI systems that truly work, what’s working (and what’s not), and why more enterprise teams are turning to custom solutions for GenAI implementation. Want to go deeper? Book a one-on-one meeting with one of our AI experts in advance. Spots are limited. Link in comments. We look forward to seeing you there! #RaiseSummit #AI21Labs #GenAI #EnterpriseAI #AIInnovation #AI Henri Delahaye | Michael Amar 🎗️| Hadrien de Cournon | Gal Badash | Joanna Kramer | Marc Taylor | Sapir Miron | Sharon Maarek | Ana Turchetti-Maia, PhD | Eytan Scemama | Jonathan Wolfson | Tal Guttman