Martin Zirulnik

Researcher, Engineer, Educator, UCLA PhD

[Martin Zirulnik @MartinZirulnik]

For more on SCP, see the paper: https://t.co/WiyKFcaDab

[Martin Zirulnik @MartinZirulnik]

Happy to share that I've joined @iSonic_ai as a Board Advisor. They build products and infrastructure for creator recognition and compensation when work is used by AI systems, and authored the Sovereign Context Protocol (SCP).

[Martin Zirulnik @MartinZirulnik]

Tens of thousands of 2025 papers contain hallucinated citations. The fix has been post-hoc verification. A protocol-level alternative: citations resolved from the retrieval envelope, not generated by the model. Fabrication becomes structurally impossible.

https://t.co/mUa4AgKZJ5

[nature @Nature]

Tens of thousands of publications from 2025 might include invalid references generated by AI, a Nature analysis suggests

https://t.co/SiIxVgJRZ7

[Martin Zirulnik @MartinZirulnik]

does this mean Claude has overcome the halting problem? (cc @AnthropicAI )

[Martin Zirulnik @MartinZirulnik]

@hwchase17 nice! 👀

[Martin Zirulnik @MartinZirulnik]

@hwchase17 Polly's counterpart that lives in your IDE / dev environment? source code (e.g. domain models, control logic, etc. that's not visible in LangSmith) can be important context for optimizing prompts.

[Martin Zirulnik @MartinZirulnik]

@OpenAI @OpenAI Read the “confessions” paper. The “self-report” is a third-person aux channel still scored by another LLM, but the “confessional” framing seems to market external critique as self-reflection. Am I wrong that this is still just LLM-as-judge?

[Martin Zirulnik @MartinZirulnik]

Huge congratulations to the amazing team at @rebindai !!🚀

Can't imagine working on a more interesting project with more talented people.

[Rebind AI @rebindai]

Exciting #BookNews! 💫@rebindai has launched with 5 timeless #classicbooks, enhanced by today’s leading thinkers. Discover #Dubliners, Thus Spoke #Zarathustra, #Walden, Heart of Darkness, and The Great #Gatsby with expert commentary. Available now. 📚🔗https://t.co/5BKBNUdyM8

[Rebind AI @rebindai]

We're honored to have acclaimed novelist John Banville guide Rebind readers through the Irish classic, Dubliners.📚Curious about how we're bringing #classicliterature to life? Join our beta. #booklovers #bookrecommendations #nyt https://t.co/h84M3MyAkA

[WIRED @WIRED]

A mysterious company called Rebind is turning notable authors into AI reading companions to create authentic conversation in hopes of making more books accessible.

And some of the top authors in the world are in support of it.

🔗 https://t.co/YEx5AvFtdm / 📷 Sahar Rana

[Martin Zirulnik @MartinZirulnik]

Great piece on @rebindai in @WIRED today! It was wonderful talking to the real @laurakipnis about what we're building—Kipnis-Bot has a lot to live up to! 🎭

[WIRED @WIRED]

Margaret Atwood, Marlon James, Lena Dunham, Roxane Gay: We’ve all agreed to be turned into AI reading companions by a mysterious company called Rebind. I report from the inside. https://t.co/gjcnInfofP

[Rebind AI @rebindai]

📰Excellent and provocative piece in @WIRED today about @rebindai from @laurakipnis, who has seen our process of bringing #classiclit to life from the inside.

[Rebind AI @rebindai]

We're honored to have #RoxaneGay teach Edith Wharton's novel The Age of Innocence, coming soon to Rebind. 📚Curious about how @rebindai is bringing #classicliterature to life? Join our waitlist.🔗http:/s.rebind.ai/x #booklovers #AI #bookrecommendations #bookclub #nyt

[Rebind AI @rebindai]

As seen in the @nytimes ! 📰Curious about how we're bringing #classicliterature to life with the help of experts like @MargaretAtwood? 📚Join the waitlist:🔗https://t.co/TMoxJePjbL #booklovers #AI #bookrecommendations #bookclub

[Martin Zirulnik @MartinZirulnik]

Congratulations to the @rebindai team on this feature! Amazing to see this work coming to life!

[Rebind AI @rebindai]

Thrilled to be featured in the @nytimes today by @skurutz ! The article highlights how our new #reading platform is not only powered by #AI and experts like @DeepakChopra and @RoxaneGay7—it’s also rooted in deep friendships and passion for #classiclit. 🔗https://t.co/S38Tj4RMTX

[Rebind AI @rebindai]

Thrilled to be featured in the @nytimes today by @skurutz ! The article highlights how our new #reading platform is not only powered by #AI and experts like @DeepakChopra and @RoxaneGay7—it’s also rooted in deep friendships and passion for #classiclit. 🔗https://t.co/S38Tj4RMTX

[Martin Zirulnik @MartinZirulnik]

Excited to finally share what we're building at @RebindAI! 📚🚀 We're turning books into conversations powered by the greatest minds and AI. Experience classic literature like never before. Join our waitlist today! https://t.co/Mnc1I46Q2P 🌐📖 #Rebind #ReadMore #InnovativeReading

[Martin Zirulnik @MartinZirulnik]

@langchain @manuelsoria_ Great piece @RLanceMartin!

[Martin Zirulnik @MartinZirulnik]

While "query construction" doesn't sound as spectacularly scifi as AI tech that makes the news, it's actually quite remarkable to watch the coevolution of natural language interfaces for databases and new ways of modeling the world with data structures.

Great piece @langchain!

[LangChain @LangChain]

⭐️ Query Construction Guide ⭐️

There's great interest in seamlessly connecting natural language with diverse data (structured, unstructured, and a mix of the two).

LLMs have impressive capacity for query construction, converting natural language into a specific query syntax for accessing each data type.

But query construction can be confusing, with several different pipelines / approaches out there.

We're releasing a guide (below) for four popular query construction pipelines summarized in one place.

We're also releasing two new related templates for:
1. Structure+unstructured data:  Text-to-SQL+semantic (w/ PostgresSQL with the Pgvector c/o @manuelsoria_)
2. Unstructured w/ metadata: Text-to-metadata filters (w/ new docs + a template for self-query retriever)

"Text-to-SQL+semantic" is an interesting recent addition to LangChain that extends "Text-to-SQL" w/ semantic queries on an embedding column.

Read more in our guide:
https://t.co/Vw5Zb4HDFu

[Martin Zirulnik @MartinZirulnik]

@DanielLSainz @langchain @Docker @neo4j @Ollama_ai I've had luck splitting html sources at the element level and adding header text as metadata to preserve the doc hierarchy. (Written a bit about this here: https://t.co/2nLccvWUP8)

Still working on this approach, and now thinking about how it might fit into a stack like this...