Rob Ó Callaghan

When it comes to AI, as California goes, so goes the nation.  In the early hours of Saturday morning, the California State Legislature passed SB 53, a bill that would require transparency from the developers of powerful, cutting-edge “frontier” AI models. SB 53 is an effort to regulate AI before the technology outpaces our attempts at oversight and the proverbial horse is well and truly out of the barn. I wrote about the debate underlying the bill, which focuses on "catastrophic risks", worst-case scenarios like AI-enabled bioweapons attacks or nuclear warfare. That framework is not without controversy, and the lines between near-term harms from deepfakes and long-term consequences aren't always clear. But the hope is that SB 53 will help prevent major disasters, and industry heavyweight Anthropic came out in support of the bill. Thank you to Thomas Woodside, Nathan Calvin, Adam Billen, Matthew Mittelsteadt, Neil Chilson, Scott S., and Irene Solaiman for speaking with me for the piece. https://lnkd.in/enm7TTX9

22

Thrilled to announce the official release of SDialog ! SDialog is an open-source Python toolkit for synthetic dialogue generation, evaluation, and analysis, designed to make creating high-quality, controllable dialogue data easier and more scalable. This toolkit has been a big part of our teamwork during JSALT workshop, and I’m really excited to see it go public! I actively participated in the development of the clinical reason for visits to inform the dialogue, the clinical dialog physical context, the clinical evaluation and the LLM as a judge which measure the clinical integrity of the generate dialogue, analyze factuality, omission, safety, communication, and cultural adaptation, and score the conversational AI from clinical depth standpoint. 👀 🔍 Key Features Key features of SDialog: - Persona-Driven Simulation: Easily create multi-agent simulations with rich personas and contexts. - Composable Orchestration: Precisely control dialogue flow and agent behavior with simple, pipe-like composition. - Built-in Evaluation: Use a suite of metrics and LLM-as-judge capabilities to evaluate and compare dialogue quality. - Mechanistic Interpretability: Go beyond black-box generation. Inspect and even steer model activations to understand and influence behavior. - LLM Interoperability: Works out-of-the-box with backends like OpenAI, HuggingFace, Ollama, AWS Bedrock, and more. I’m very proud of this release, and especially of the clinical evaluation tools. 🔗 Check it out on GitHub (Idiap Research Institute): https://lnkd.in/eJwiXNZb Give it a ⭐️ if you find it interesting! A big thanks to Sergio Burdisso, Yanis Labrak, Séverin BAROUDI, Thomas Schaaf, Andrew Perrault, Amy Chun, and all the others. Also, special thanks to Ricard Marxer & Sanjeev Khudanpur for giving me the opportunity to participate to JSALT25 !

28

Some fundamental limitations of LLMs: * constrained context window * poor integration with real systems * poor long-horizon task execution I found this video enlightening on how AI agents overcome these limitations and let LLMs shine in broad, real-world applications.

1

There's a fundamental problem with the interaction between politics and Open Source which is that politicians don't understand that a software "vendor" can just be some person in their spare time with a Github account. They think of computers and software as "products" that are produced by megacorporations when the reality is that software is malleable and something you can create easily. Understanding that software is completely under your control is a superpower both in your everyday life and in research and it's one of the realities that I try desperately to get students to understand.

11

1 Comment

this is not AI. no gradient descent. no loss function. no model that learned to make a fly walk and eat. eon systems took the existing adult drosophila connectome - 140,000 neurons, ~50 million connections - and ran it in a simulation. directly. the virtual fly navigates toward food, grooms itself when virtual dust accumulates, forages. those behaviors aren't programmed. they emerge from the wiring. the connectome is just a graph of neurons and the connections between them. what eon showed is that if you simulate that graph with basic integrate-and-fire dynamics and attach it to a physics-simulated body, recognizable behavior comes out the other side. we've spent decades asking what the wiring diagram encodes, before you add anything else. for a fruit fly, apparently: hunger, taste, grooming, navigation. a lot. the human connectome project has been running since 2009. we don't have a complete wiring diagram yet, but progress is being made. we hope this succesful result accelerates that progress. wild future ahead.

364

21 Comments