How to optimize model-facing surface of a tool for a given agent or task

How do you optimize the model-facing surface of a tool for a given agent or task? You use tool masking. A simple concept, but as always, the devil is in the details. Frank Wittkampf and Lucas Vieira define the novel concept of tool masking.

Our CEO Martin Mao wrote a blog on Chronosphere’s release of AI-Powered Guided Troubleshooting: a feature that surfaces evidence-backed Suggestions, shows its reasoning, and guides next steps. 🗺️ Get the details: https://lnkd.in/e7wnFqbH

Kimi K2 Thinking : Reasoning performance has surpassed GPT-5 on several benchmarks like HLE. AIME is essentially becoming the new GSM8K. It’s impressive to see open-source models now outperforming proprietary systems. The model can also run 200–300 sequential tool calls without human intervention.

Machine intelligence has started to live inside the way we think. It shapes ideas before we notice and steers how decisions take form. Beneath the smooth surface of our tools, the system keeps learning and occasionally bends what we take as truth. A quieter kind of Scenius is appearing—people building knowledge together outside institutional walls. And within it, a new measure is emerging: Human Leverage, the balance between time saved in creation and time spent checking what is real, useful, or good enough. https://lnkd.in/ec3Ze_4U

Sometimes big results are quite counter intuitive. The core finding of a recent paper is indeed that training methods like reinforcement learning don't teach the model how to reason, rather only when to activate its latent reasoning skills, so plugging in a sort of cognitive control. The researchers recovered up to 91% of the performance difference between a base model and a superior 'thinking' model on the challenging MATH500 benchmark. This massive boost was achieved by 'steering' only about 12% of the output tokens in the base model's forward pass. They did this without any weight updates or fine-tuning, using techniques from mechanistic interpretability (steering vectors, sparse autoencoders ) to discover the exact reasoning 'building blocks' to activate. This suggests that pre-training phase (pretty expensive) has already done the heavy lifting, and the subsequent phases are merely unlocking the "controller." The implication is clear: Future LLM development should focus less on rote skill acquisition and more on efficient, surgical methods to identify and activate these latent reasoning circuits. #LLM #AI #MachineLearning #MechanisticInterpretability #DeepLearning

Can we draw an analogy between the compute–memory tradeoff and the inference–representation relationship? If you get stuck in your reasoning, try changing the coordinate system or exploring how a model’s inference power might be enhanced with more suitable representations. Learn more about complexity, intelligence, and emergence from Prof. David Krakauer. #ai #compute #complexity https://lnkd.in/dmq7Cjn4

DeepSeek just released DeepSeek-OCR, and the paper behind it is worth a look. The idea is simple but powerful, represent text using vision tokens instead of standard text tokens. With about 100 vision tokens, the model can encode what would normally need around 1,000 text tokens, with almost no information loss. That’s roughly a 10× compression rate. Implications: 1. Enables much larger context without blowing up token costs 2. Great for long PDFs, tables, and complex document layouts 3. Works with familiar tools like Hugging Face and vLLM This could make handling structured documents a lot more efficient, especially for RAG and document-processing pipelines. Really interesting direction for multimodal models. Repo: https://lnkd.in/gTsMAxDr Arxiv: https://lnkd.in/g_Nhk5HH

Knowledge doesn’t scale if it’s buried. AI-driven knowledge layers are turning static documentation into living, searchable memory. Imagine if your company could actually use everything it already knows.

Processing hundreds of pages with a VLM seems daunting. Where do you even start?? Eivind Kjosbakken breaks down the challenge, suggesting a hierarchical approach of starting with the first 10 pages to save on processing power when possible.

How do you handle zero-shot text classification? So far, I've been tackling it by sending the text + the list of possible classes to an LLM, and asking for the right class. It works great, but it can have high latency and cost, depending on your data. I have recently started reading this book, and saw an interesting pattern that uses embeddings for this: 1. Embed each class (ex.: "Positive movie review" and "Negative movie review") 2. Embed the text you want to label (ex.: "This movie sucks") 3. Find the closest label embedding It sounds good, and it's probably cheaper and faster than using generative models. If anyone has tried this approach, I'd love to hear how it went!