NJM Insurance Group

🤔 Thinking Machines Lab 's just released Interaction Models 🔗 https://lnkd.in/gFtMcEsi 🔹 Most AI systems are still fundamentally turn-based: the user provides an input, the model processes it, and only then returns an output. 🔹 Thinky proposes a different abstraction called interaction models, where the model continuously takes in audio, video, and text while simultaneously thinking, responding, and acting in real time. 🔹 This changes the interface from request-response to full-duplex interaction. Current models assume a strict boundary between context ingestion and token generation. Once decoding begins, the model typically cannot incorporate new information until the next turn. 🔹 Interaction models remove this boundary by allowing perception, reasoning, and generation to proceed concurrently. 🔹 Rather than operating over a static context window, the model maintains a continuously updated latent state that integrates incoming signals while producing outputs. Inference becomes a streaming process in which context is no longer fixed at the start of generation, but evolves throughout the interaction. 💡 This has important consequences for systems design. - Input and output become asynchronous streams - State must be updated incrementally rather than rebuilt each turn - The model can revise its response mid-generation - Reasoning remains grounded in a changing environment - The result is a more natural form of collaboration. 🔹 A model can translate speech while listening, interrupt itself when the user changes direction, respond to visual events as they occur, or adapt continuously instead of waiting for the next prompt. Thinking Machines describes this as addressing a bandwidth bottleneck between humans and AI, where much of a user's intent and judgment is lost when communication is restricted to discrete turns. 🔹 If prompt engineering shaped how we write instructions and context engineering shaped how we manage model state, interaction models may shape how models and humans collaborate in real time.

Bring Your Child to Work Day at Crum & Forster ! Originally launched as “Take Our Daughters to Work Day” in the early ’90s, the day was created to inspire the next generation and give kids a window into the working world. Today, it’s about exposure, confidence, and showing what’s possible. From hands-on activities to meaningful projects, the kids spent the day supporting local organizations—including making dog toys for St. Hubert’s Animal Welfare Center—a small act that goes a long way for animals waiting for a home. A great reminder that doing good and doing well can go hand in hand. (And I’m told the 3PM ice cream stop was a highlight.) Thank you, Crum & Forster for a great day!

🔄 Looped Transformer 🔗 https://lnkd.in/g4KAFFJN With the release of Anthropic 's Claude Mythos, there has been a lot of speculation that a looped / recurrent transformer variant might be part of the architecture (and projects like OpenMythos back this up in a concrete way: https://lnkd.in/gRHcSzRv ). Whether that’s true or not, the underlying idea is worth paying attention to. One of the key result in “Looped Transformers as Programmable Computers” by Angeliki Giannou , Shashank Rajput , Jy-yong Sohn , Kangwook Lee , Jason D. Lee, and Dimitris Papailiopoulos is that you can decouple model depth from computation length by introducing an external loop over a fixed transformer. Instead of stacking layers to simulate multi-step reasoning, you reuse the same network repeatedly and treat the sequence as mutable state. The input is structured explicitly: a scratchpad (working space), a memory region, and a command segment. Positional encodings are not just for ordering, they act as pointers for addressing memory and instructions. What’s actually interesting is how standard transformer components are repurposed into low-level compute primitives: 🧠 Attention is engineered to behave like a soft routing mechanism, enabling precise read/write operations between memory and scratchpad (effectively approximate permutation matrices under high temperature softmax). ⚙️ Feedforward layers implement bit-level logic using ReLU, enabling operations like integer addition, negation, and program counter updates. 🔀 Conditional branching is constructed by extracting sign information from stored values and updating a program counter accordingly. These primitives compose into a SUBLEQ-style one-instruction set computer inside a transformer, meaning the model can execute arbitrary programs encoded in the input. The paper extends this to a more general instruction (FLEQ) where the “operation” can be any function block (e.g. matrix ops, nonlinear transforms) hardcoded into the weights. The important takeaway is not that transformers are Turing complete (simulate a Turing machine), but that you can realize this with constant depth and push all iterative structure into recurrence over the sequence. That reframes how to think about reasoning in LLMs: not as deeper stacks of layers, but as fixed compute blocks executing implicit programs over evolving state.