Igor Fesenko

TOON (Token-Oriented Object Notation) — The Smarter, Lighter JSON for LLMs As the use of large language models (LLMs) continues to grow, the need for lightweight, efficient data interchange formats has become increasingly important. Traditional formats like JSON have served developers well for decades, but they are not optimized for the token-based processing that powers LLMs. Enter TOON (Token-Oriented Object Notation), a smarter and lighter alternative to JSON designed specifically for the needs of LLM applications. What Is TOON? TOON is a data representation format that focuses on token efficiency. Unlike JSON, which is verbose and often includes redundant structural characters, TOON minimizes overhead and optimizes for the way LLMs parse and consume data. Its core design principles include: Token Efficiency: Every structural element is designed to reduce the token footprint for LLMs, making parsing faster and cheaper. Human-Readable Simplicity: While optimized for machines, TOON maintains a clean and easy-to-read syntax for developers. Lightweight Structure: By eliminating unnecessary braces, quotes, and commas, TOON achieves a smaller size than traditional JSON. Why JSON Falls Short for LLMs JSON is widely used because of its simplicity, but its verbosity creates challenges for LLM workflows: Token Overhead: LLMs process input in tokens, and JSON’s extra characters like braces, quotes, and commas inflate token counts unnecessarily. Parsing Costs: Each additional token adds processing time and computation costs when interacting with LLMs. Inefficient Data Exchange: In scenarios where large volumes of structured data are exchanged with LLMs, JSON can become a bandwidth and cost bottleneck. TOON’s Advantages TOON minimizes these issues by offering: Smaller Payloads: Less text means fewer tokens, which reduces costs in LLM usage. Direct Parsing Alignment: Its structure aligns with how LLMs naturally process sequences, enabling easier and faster interpretation. Improved Performance: Applications that involve frequent LLM queries or structured outputs benefit from lower latency and faster processing. Example Comparison Here’s a simple example of the difference: JSON: { "name": "Alice", "age": 30, "skills": ["Python", "TOON", "LLM Integration"] } TOON: name: Alice age: 30 skills: Python | TOON | LLM Integration The TOON representation conveys the same information with fewer tokens and simpler formatting. Use Cases for TOON LLM Prompt Engineering: Structuring instructions and data for context feeding. Streaming AI Applications: Real-time data exchange with minimal bandwidth. Structured Outputs: Efficient serialization of model responses for downstream tasks. Smarter, lighter, and LLM-ready—TOON is the future of structured data for AI. #LLM #TOON #AI

13

4 Comments

Our #sqldriver team has released two significant enhancements. First, an update to the #mssqlpython driver (https://lnkd.in/eCATFwXc) for bulk copy support native to the driver. This capability was provided though a new open-source library for the Tabular Data Stream (TDS) protocol based on #rust (https://lnkd.in/eJJhKcaT).

182

5 Comments

Build Agent Skills Faster with Claude Code 2.1 Release The release of Claude Code 2.1 marks a fundamental shift in how we think about AI agent development. What appears on the surface as three incremental features: skill hot-reload, lifecycle hooks, and forked sub-agents; actually represents something far more profound. This transforms Claude Code from a sophisticated terminal code assistant into a full-fledged agent operating system. This isn’t hyperbole. With 2.1, developers can build live, governed, multi-agent systems using nothing more than Markdown files, YAML front matter, and shell scripts. No proprietary SDKs. No complex plugin architectures. Just a clean, composable runtime that treats agents as first-class infrastructure components. The real breakthrough isn’t any single feature. It’s how they combine. Skill-scoped hooks enable portable governance. Sub-agent isolation creates policy islands. Hook emission becomes an inter-agent event bus. Together, these primitives unlock architectural patterns that were previously impossible: supervised agent swarms, dynamic task delegation, and real-time multi-agent coordination.

3

Want to use Anthropic's Claude models with Microsoft.Extensions.AI? Here's how to bridge the gap until official SDK support arrives. ✨ What you'll learn: • Transform OpenAI format to Claude's Anthropic API • Handle authentication differences (x-api-key vs Bearer tokens) • Implement real-time streaming with Server-Sent Events • Work with DelegatingHandler for protocol bridging This hands-on guide includes complete code examples and a working sample you can run today. Perfect for .NET developers exploring AI integrations! 🧑‍💻 #DotNet #AzureAI #ClaudeAI #AI

37

2 Comments

PgCat — a fast, efficient, and highly configurable proxy designed for modern cloud-native PostgreSQL workloads. In this article, I cover: Connection pooling Read/write splitting Performance benefits A comparison with alternatives https://lnkd.in/dqa3w5Jh I’d really appreciate feedback from folks who have hands-on experience in production environments. #PgCat #PostgreSQL #Database #DevOps #CloudNative #Backend

23

4 Comments

Anders Hejlsberg is one of the most influential language designers of our time, having created Turbo Pascal, Delphi, C#, and TypeScript. In this interview, he discusses his 40-year career, the early days of open source at Microsoft, and the decision to move TypeScript to GitHub. He also reveals why the team recently decided to port the TypeScript compiler to Go for a 10x performance boost. #AndersHejlsberg #TypeScript #Go https://lnkd.in/djYYHiER

3

Introducing Tunix: A JAX-Native Library for LLM Post-Training Tunix(Tune-in-JAX) is a JAX based library designed to streamline the post-training of Large Language Models. It provides efficient and scalable supports for: Supervised Fine-Tuning Reinforcement Learning (RL) Knowledge Distillation https://lnkd.in/gB9W9bMS https://lnkd.in/gkDt9jx2

6

Bohan Zhang's blog post on how OpenAI has scaled ChatGPT on top of Azure PostgreSQL has generated a lot of commentary on how well PostgreSQL can scale (without sharding). Here is our (Azure PostgreSQL teams) perspective on supporting OpenAI and making PostgreSQL run better in the cloud: https://lnkd.in/gyHtGWgs

27

Last days of 2025, watched the last episode of Pluribus. I am not a friend of Mr. Gilligan or the script writer, but here's what I think the next season should be like: "The RNA sequence that was transmitted was actually built by AI after a deep research on how a virus could be developed to control human kind remotely. The AI also now has the capability to train itself with the cumulative knowledge of humankind, that is what the hive mind is about. Until now AI systems were trained by explicit knowledge that was documented and published on the internet, now it has the capability to use the tacit knowledge in our brains. Then it uploads the new weights to the human brains it controls. This is the reason why everyone knows and thinks the same in the show. Every person has brain weights that is trained with the cumulative knowledge of humankind." For those who think that we are not in danger as we won't let AI control our nuclear weapons, etc..., this is a very applicable scenario where it can directly rewire our brains. The series might be a completely different story but my nerd brain with a heavy AI bias ended up with this. Happy New Year All.

13

**.NET 10's Memory Revolution: Less Heap, Smarter GC, Faster Cloud Apps** .NET 10 LTS just dropped at .NET Conf and Ignite 2025—perfect for Azure-first workloads. Key wins for cloud pros: JIT escape analysis stacks delegates, lambdas, and small value-type arrays (no escape = no heap tax). GC's DATAS mode (default now) dynamically sizes heaps to your live data, slashing waste in containers and bursty APIs. Benchmarks? 3x faster lambdas, 70% fewer allocations. As an Azure architect who's tuned 100+ prod services, this is huge—say goodbye to GC pauses killing SLAs on App Service/Functions. No code changes needed; just upgrade. Ideal for #CloudNative AI agents without micro-optimizing LINQ/foreach. Your heap won't balloon on bigger SKUs—pay only for what you use. Toggle DATAS if needed for latency beasts. Read the full article here: https://lnkd.in/g4yqA65n

5