Tansu, an open-source, Apache Kafka-compatible messaging broker, challenges traditional approaches by prioritizing statelessness. Instead of replicating data like Kafka, Tansu delegates durability to external storage, allowing for brokers that are lightweight ("cattle," not "pets") and scale rapidly. It supports various storage backends like S3, SQLite, and Postgres, with a particular emphasis on Postgres integration for streamlined data pipelines. Tansu also offers broker-side schema validation and the ability to directly write validated data to open table formats like Iceberg, Delta Lake, or Parquet. The project is written in Rust and seeks contributors.
Júlio Falbo argues that integrating AI into engineering organizations is hampered by complex connection methods, proposing a solution centered around “SKILL.md” – Markdown files defining tool usage – and “AI Gateways” for centralized orchestration. This combination fosters an “AI-native architecture” prioritizing ease of use, governance, and scalability over bespoke integrations. Ultimately, this approach shifts the focus from complex coding to clear documentation, democratizing AI tool access and boosting productivity.
* Simplifies AI integration via Markdown-based "skills."
* Utilizes AI Gateways for centralized control and security.
* Promotes a convention-over-configuration approach for AI systems.
* Simplifies AI integration via Markdown-based "skills."
* Utilizes AI Gateways for centralized control and security.
* Promotes a convention-over-configuration approach for AI systems.
Developers are replacing bloated MCP servers with Markdown skill files — cutting token costs by 100x. This article explores a two-layer architecture emerging in production AI systems, separating knowledge from execution. It details how skills (Markdown files) encode stable knowledge, while MCP servers handle runtime API interactions. The piece advocates for a layered approach to optimize context window usage, reduce costs, and improve agent reasoning by prioritizing knowledge representation in a version-controlled, accessible format.
This article explains the differences between Model Context Protocol (MCP), Retrieval-Augmented Generation (RAG), and AI Agents, highlighting that they solve different problems at different layers of the AI stack. It also covers how ChatGPT routes prompts and handles modes, agent skills, architectural concepts for developers, and service deployment strategies.
This article details research into finding the optimal architecture for small language models (70M parameters), exploring depth-width tradeoffs, comparing different architectures, and introducing Dhara-70M, a diffusion model offering 3.8x faster throughput with improved factuality.
LLMs are powerful for understanding user input and generating human‑like text, but they are not reliable arbiters of logic. A production‑grade system should:
- Isolate the LLM to language tasks only.
- Put all business rules and tool orchestration in deterministic code.
- Validate every step with automated tests and logging.
- Prefer local models for sensitive domains like healthcare.
| **Issue** | **What users observed** | **Common solutions** |
|-----------|------------------------|----------------------|
| **Hallucinations & false assumptions** | LLMs often answer without calling the required tool, e.g., claiming a doctor is unavailable when the calendar shows otherwise. | Move decision‑making out of the model. Let the code decide and use the LLM only for phrasing or clarification. |
| **Inconsistent tool usage** | Models agree to user requests, then later report the opposite (e.g., confirming an appointment but actually scheduling none). | Enforce deterministic tool calls first, then let the LLM format the result. Use “always‑call‑tool‑first” guards in the prompt. |
| **Privacy concerns** | Sending patient data to cloud APIs is risky. | Prefer self‑hosted/local models (e.g., LLaMA, Qwen) or keep all data on‑premises. |
| **Prompt brittleness** | Adding more rules can make prompts unstable; models still improvise. | Keep prompts short, give concrete examples, and test with a structured evaluation pipeline. |
| **Evaluation & monitoring** | Without systematic “evals,” failures go unnoticed. | Build automated test suites (e.g., with LangChain, LangGraph, or custom eval scripts) that verify correct tool calls and output formats. |
| **Workflow design** | Treat the LLM as a *translator* rather than a *decision engine*. | • Extract intent → produce a JSON/action spec → execute deterministic code → have the LLM produce a user‑friendly response. <br>• Cache common replies to avoid unnecessary model calls. |
| **Alternative UI** | Many suggest a simple button‑driven interface for scheduling. | Use the LLM only for natural‑language front‑end; the back‑end remains a conventional, rule‑based system. |
- Isolate the LLM to language tasks only.
- Put all business rules and tool orchestration in deterministic code.
- Validate every step with automated tests and logging.
- Prefer local models for sensitive domains like healthcare.
| **Issue** | **What users observed** | **Common solutions** |
|-----------|------------------------|----------------------|
| **Hallucinations & false assumptions** | LLMs often answer without calling the required tool, e.g., claiming a doctor is unavailable when the calendar shows otherwise. | Move decision‑making out of the model. Let the code decide and use the LLM only for phrasing or clarification. |
| **Inconsistent tool usage** | Models agree to user requests, then later report the opposite (e.g., confirming an appointment but actually scheduling none). | Enforce deterministic tool calls first, then let the LLM format the result. Use “always‑call‑tool‑first” guards in the prompt. |
| **Privacy concerns** | Sending patient data to cloud APIs is risky. | Prefer self‑hosted/local models (e.g., LLaMA, Qwen) or keep all data on‑premises. |
| **Prompt brittleness** | Adding more rules can make prompts unstable; models still improvise. | Keep prompts short, give concrete examples, and test with a structured evaluation pipeline. |
| **Evaluation & monitoring** | Without systematic “evals,” failures go unnoticed. | Build automated test suites (e.g., with LangChain, LangGraph, or custom eval scripts) that verify correct tool calls and output formats. |
| **Workflow design** | Treat the LLM as a *translator* rather than a *decision engine*. | • Extract intent → produce a JSON/action spec → execute deterministic code → have the LLM produce a user‑friendly response. <br>• Cache common replies to avoid unnecessary model calls. |
| **Alternative UI** | Many suggest a simple button‑driven interface for scheduling. | Use the LLM only for natural‑language front‑end; the back‑end remains a conventional, rule‑based system. |
The article provides practical advice for software architects on how to effectively communicate and deploy ideas through documentation. Key takeaways include:
1. **Focus on ideas, not code**: Architects must organize and deploy ideas to people, not just machines.
2. **Use bullet points**: They help structure information clearly and make documents easy to skim.
3. **Structure with headers**: Break content into sections for easy navigation and quick information retrieval.
4. **Write for the reader**: Prioritize clarity and relevance over perfect formatting or templates.
5. **Organize chronologically**: Group documents by time (year/sprint) rather than topic to improve searchability.
6. **Document types matter**: Specific document formats like architecture overviews, dev designs, and project proposals help manage complex projects.
7. **Keep documents concise and useful**: Aim for point-in-time documentation that remains useful even if outdated.
8. **Share and iterate**: Distribute documents widely and seek feedback to improve them.
1. **Focus on ideas, not code**: Architects must organize and deploy ideas to people, not just machines.
2. **Use bullet points**: They help structure information clearly and make documents easy to skim.
3. **Structure with headers**: Break content into sections for easy navigation and quick information retrieval.
4. **Write for the reader**: Prioritize clarity and relevance over perfect formatting or templates.
5. **Organize chronologically**: Group documents by time (year/sprint) rather than topic to improve searchability.
6. **Document types matter**: Specific document formats like architecture overviews, dev designs, and project proposals help manage complex projects.
7. **Keep documents concise and useful**: Aim for point-in-time documentation that remains useful even if outdated.
8. **Share and iterate**: Distribute documents widely and seek feedback to improve them.
Sam Newman discusses the three golden rules of distributed computing and how they necessitate robust handling of timeouts, retries, and idempotency. He provides practical, data-driven strategies for implementing these principles, including using request IDs and server-side fingerprinting to create safe, resilient distributed systems.
This article explores the construction and evolution of ancient Greek temples, highlighting the three classical column styles – Doric, Ionic, and Corinthian – noting that Corinthian columns originated in Roman civilization. It details the progression from early mud brick structures to the enduring stone temples, exemplified by sites like Temple C in Selinus, Sicily, and the Temple of Apollo at Didyma, Turkey. The piece emphasizes the Greeks’ innovative use of columns, often inspired by sacred forests, and references related content showcasing reconstructions and replicas of ancient Greek
A detailed comparison of the architectures of recent large language models (LLMs) including DeepSeek-V3, OLMo 2, Gemma 3, Mistral Small 3.1, Llama 4, Qwen3, SmolLM3, and Kimi 2, focusing on key design choices and their impact on performance and efficiency.
1. **DeepSeek V3/R1**:
- Uses Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE) for efficiency.
- MLA compresses key and value tensors to reduce KV cache memory usage.
- MoE activates only a subset of experts per token, improving inference efficiency.
2. **OLMo 2**:
- Focuses on transparency in training data and code.
- Uses RMSNorm layers placed after attention and feed-forward modules (Post-Norm).
- Introduces QK-Norm, an additional RMSNorm layer applied to queries and keys inside the attention mechanism.
3. **Gemma 3**:
- Employs sliding window attention to reduce memory requirements in the KV cache.
- Uses a 5:1 ratio of sliding window attention to global attention layers.
- Combines Pre-Norm and Post-Norm RMSNorm layers around the attention module.
4. **Mistral Small 3.1**:
- Outperforms Gemma 3 27B on several benchmarks while being faster.
- Uses a standard architecture with a custom tokenizer and reduced KV cache and layer count.
5. **Llama 4**:
- Adopts an MoE approach similar to DeepSeek V3 but with fewer, larger experts.
- Alternates MoE and dense modules in every other transformer block.
6. **Qwen3**:
- Comes in both dense and MoE variants.
- Dense models are easier to fine-tune and deploy, while MoE models are optimized for scaling inference.
7. **SmolLM3**:
- Uses No Positional Embeddings (NoPE), omitting explicit positional information injection.
- NoPE improves length generalization, meaning performance deteriorates less with increased sequence length.
8. **Kimi K2 and Kimi K2 Thinking**:
- Uses a variant of the Muon optimizer over AdamW.
- Kimi K2 Thinking extends the context size to 256k tokens.
9. **GPT-OSS**:
- OpenAI's first open-weight models since GPT-2.
- Uses sliding window attention and a width-versus-depth trade-off.
10. **Grok 2.5**:
- Uses a small number of large experts and a shared expert module.
- Reflects an older trend in MoE architectures.
11. **GLM-4.5**:
- Comes in two variants: a 355-billion-parameter model and a more compact 106-billion-parameter version.
- Uses a shared expert and starts with several dense layers before introducing MoE blocks.
12. **Qwen3-Next**:
- Introduces a Gated DeltaNet + Gated Attention hybrid mechanism.
- Uses Multi-Token Prediction (MTP) for efficiency.
13. **MiniMax-M2**:
- Uses per-layer QK-Norm and partial RoPE.
- More "sparse" than Qwen3, with fewer active experts per token.
14. **Kimi Linear**:
- Modifies the linear attention mechanism with Kimi Delta Attention (KDA).
- Combines Gated DeltaNet with Multi-Head Latent Attention (MLA).
15. **Olmo 3 Thinking**:
- Uses sliding window attention and YaRN for context extension.
- Comes in base, instruct, and reasoning variants.
16. **DeepSeek V3.2**:
- Adds a sparse attention mechanism to improve efficiency.
- On par with GPT-5.1 and Gemini 3.0 Pro on certain benchmarks.
17. **Mistral 3**:
- First MoE model since Mixtral in 2023.
- Partnered with NVIDIA for optimization on Blackwell chips.
18. **Nemotron 3**:
- A Transformer-Mamba hybrid architecture.
- Interleaves Mamba-2 sequence-modeling blocks with sparse MoE feed-forward layers.
19. **Xiaomi MiMo-V2-Flash**:
- Uses sliding window attention in a 5:1 ratio with global attention.
- Employs multi-token prediction (MTP) for efficiency.
20. **Arcee AI Trinity Large**:
- Uses alternating local:global attention layers, NoPE, and gated attention.
- Introduces depth-scaled sandwich norm for training stability.
1. **DeepSeek V3/R1**:
- Uses Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE) for efficiency.
- MLA compresses key and value tensors to reduce KV cache memory usage.
- MoE activates only a subset of experts per token, improving inference efficiency.
2. **OLMo 2**:
- Focuses on transparency in training data and code.
- Uses RMSNorm layers placed after attention and feed-forward modules (Post-Norm).
- Introduces QK-Norm, an additional RMSNorm layer applied to queries and keys inside the attention mechanism.
3. **Gemma 3**:
- Employs sliding window attention to reduce memory requirements in the KV cache.
- Uses a 5:1 ratio of sliding window attention to global attention layers.
- Combines Pre-Norm and Post-Norm RMSNorm layers around the attention module.
4. **Mistral Small 3.1**:
- Outperforms Gemma 3 27B on several benchmarks while being faster.
- Uses a standard architecture with a custom tokenizer and reduced KV cache and layer count.
5. **Llama 4**:
- Adopts an MoE approach similar to DeepSeek V3 but with fewer, larger experts.
- Alternates MoE and dense modules in every other transformer block.
6. **Qwen3**:
- Comes in both dense and MoE variants.
- Dense models are easier to fine-tune and deploy, while MoE models are optimized for scaling inference.
7. **SmolLM3**:
- Uses No Positional Embeddings (NoPE), omitting explicit positional information injection.
- NoPE improves length generalization, meaning performance deteriorates less with increased sequence length.
8. **Kimi K2 and Kimi K2 Thinking**:
- Uses a variant of the Muon optimizer over AdamW.
- Kimi K2 Thinking extends the context size to 256k tokens.
9. **GPT-OSS**:
- OpenAI's first open-weight models since GPT-2.
- Uses sliding window attention and a width-versus-depth trade-off.
10. **Grok 2.5**:
- Uses a small number of large experts and a shared expert module.
- Reflects an older trend in MoE architectures.
11. **GLM-4.5**:
- Comes in two variants: a 355-billion-parameter model and a more compact 106-billion-parameter version.
- Uses a shared expert and starts with several dense layers before introducing MoE blocks.
12. **Qwen3-Next**:
- Introduces a Gated DeltaNet + Gated Attention hybrid mechanism.
- Uses Multi-Token Prediction (MTP) for efficiency.
13. **MiniMax-M2**:
- Uses per-layer QK-Norm and partial RoPE.
- More "sparse" than Qwen3, with fewer active experts per token.
14. **Kimi Linear**:
- Modifies the linear attention mechanism with Kimi Delta Attention (KDA).
- Combines Gated DeltaNet with Multi-Head Latent Attention (MLA).
15. **Olmo 3 Thinking**:
- Uses sliding window attention and YaRN for context extension.
- Comes in base, instruct, and reasoning variants.
16. **DeepSeek V3.2**:
- Adds a sparse attention mechanism to improve efficiency.
- On par with GPT-5.1 and Gemini 3.0 Pro on certain benchmarks.
17. **Mistral 3**:
- First MoE model since Mixtral in 2023.
- Partnered with NVIDIA for optimization on Blackwell chips.
18. **Nemotron 3**:
- A Transformer-Mamba hybrid architecture.
- Interleaves Mamba-2 sequence-modeling blocks with sparse MoE feed-forward layers.
19. **Xiaomi MiMo-V2-Flash**:
- Uses sliding window attention in a 5:1 ratio with global attention.
- Employs multi-token prediction (MTP) for efficiency.
20. **Arcee AI Trinity Large**:
- Uses alternating local:global attention layers, NoPE, and gated attention.
- Introduces depth-scaled sandwich norm for training stability.
