🥑 Yet Another (AI) Yuval

Join @YuvalinTheDeep, Senior Developer Advocate at @AI21Labs, for a live webinar in partnership with @DataCamp: The Four Gaps Between Demo Agents and Production Systems. If you're shipping agents to prod, this is where prototypes break and real systems begin. 🗓 May 6 · 11am ET Register:datacamp.com/webinars/the-f…

Day 1 of #AIDevSF is live. 🙌 📍 Find us at booth 121 🎤 Catch our Chief Product Officer Or Dagan's session on Day 2: Efficiently Navigating the Agentic Action Space: Meta-Model Orchestration using AI21 Maestro. Thanks @DeepLearningAI for having us.

4/ Full breakdown + the conversation with Barak: ai21.com/blog/mind-the-…

3/ None of these are exotic research ideas. They're just what production systems actually need to do. The gap between a demo and something that runs reliably at 2am isn't the model. The good news: we know exactly how to build it.

1/ "Agent" is becoming a meaningless word 🤨 Every LLM pipeline is suddenly "agentic." Every wrapper is an "agent framework." The word has been stretched until it broke.

Here's Eran explaining this + Adding a new layer to MCP: ai21.com/yaap/mcp-was-b…

Parallel writes are becoming the real bottleneck, and it's nice to see it in CC

@bcherny how we at @AI21Labs structured “workspace” primitives around this pattern: ai21.com/blog/stateful-…

@bcherny +1, parallel writes are the real bottleneck, and I’m glad this landed in the CLI now. Worktrees give each agent an isolated workspace so you can compare diffs, merge the winner, and delete the rest. We recently wrote a technical blog about >>

Introducing: built-in git worktree support for Claude Code Now, agents can run in parallel without interfering with one other. Each agent gets its own worktree and can work independently. The Claude Code Desktop app has had built-in support for worktrees for a while, and now we're bringing it to CLI too. Learn more about worktrees: git-scm.com/docs/git-workt…

Full writeup with experiments, qualitative examples, and implementation details (including Jupyter notebook + Seinfeld dataset): ai21.com/blog/query-dep…

Results on MTEB: 7/8 configs improved. We also got significant improvements on FinanceBench, NarrativeQA, QMSUM and our Seinfeld datasets. Tradeoff: You're storing the corpus K times (K = number of chunk sizes) For most applications, storage cost < retrieval quality gains >>

Tuning the chunking phase in RAG systems is notoriously difficult, and the feeling is that there's no "optimal" chunking method. Can it be that different queries just need different chunk sizes, even over the same corpus? (Work by @AI21Labs) >>

📣 Calling founders and builders interested in AI productivity agents: Productivity Agent Build Day Hackathon this Saturday on September 20, 10 AM PT at AGI House 👉 RSVP here to secure your spot: app.agihouse.org/events/product…

@profcelsofontes @AI21Labs working on it 😎

@AI21Labs Such a pity jamba 1.7 is not avaliable on openrouter

Attention was never enough. The hybrid LLM era is here—and it’s moving fast. From Mamba to Jamba to Bamba, we mapped every major model that’s challenged the Transformer default in the past 18 months. 🧵 A timeline of what’s changed and why it matters ↓ 🔗 ai21.com/blog/rise-of-h…

@Whatevercrypto @AI21Labs

@AI21Labs None of them seemed to have stuck though, unless one of the closed labs is using one of these architectures without telling us

@AaryanBhagat13 @AI21Labs Maybe in the future :)

@AI21Labs You have done an analysis of State Space Models and its variants I believe. But have you considered writing a similar blogpost about Diffusion LLMs (on text)?