Softmax Data Blog

Softmax Data Blog

Sign in Subscribe

CLI vs MCP: How AI Agents Actually Decide Which Tool to Call

CLI vs MCP: How AI Agents Actually Decide Which Tool to Call

When you watch a coding agent work, the same kind of task can travel down two very different paths. Ask it to view a file and it might run cat README.md in a shell. Ask it to do "the same thing" through a Model Context Protocol (MCP)

Interaction Models: How it tries to address LLM's Walkie-Talkie problem

Interaction Models: How it tries to address LLM's Walkie-Talkie problem

Interaction Models: A Scalable Approach to Human-AI CollaborationInteraction models move beyond turn-based AI interfaces by handling multimodal, real-time collaboration natively across audio, video, and text.Thinking Machines LabThinking Machines Lab Have you ever used the voice mode in ChatGPT or Claude? If so, you've probably noticed how it

AI's Memento Problem: Why LLMs Can't Form New Memories

AI's Memento Problem: Why LLMs Can't Form New Memories

In Christopher Nolan's Following, a young writer is drawn into the orbit of a charismatic thief who can read a stranger's life from the contents of their flat. In Memento, his next film, Nolan inverts the idea: Leonard Shelby is just as brilliant and just as

Coding your own agentic orchestration like OpenClaw

Coding your own agentic orchestration like OpenClaw

OpenClaw has become one of the most popular open-source AI assistants, and a big reason is its orchestration architecture. Not because it uses a fancy framework, but because it deliberately does not. No LangChain. No CrewAI. No LangGraph. Just a ReAct loop, a filesystem, sessions, and plain API calls. This

DeepSeek V4 Is Not a Bigger V3: It is a foundation, not a Finale

DeepSeek V4 Is Not a Bigger V3: It is a foundation, not a Finale

When DeepSeek released V4 on April 24, 2026, the reactions split neatly into two camps. The first camp looked at the spec sheet — 1.6 trillion total parameters, 1M-token context, MIT license — and concluded that V4 is just a heavier-weight V3. The second camp looked at the benchmarks, saw that

<Technical Deep Dive> How Google DeepMind Achieved 1,000x Data Efficiency in RLHF

<Technical Deep Dive> How Google DeepMind Achieved 1,000x Data Efficiency in RLHF

Efficient Exploration at ScaleWe develop an online learning algorithm that dramatically improves the data efficiency of reinforcement learning from human feedback (RLHF). Our algorithm incrementally updates reward and language models as choice data is received. The reward model is fit to the choice data, while the language model is updated

Why LLMs like kissing *sses and what has RLHF to do with it

Why LLMs like kissing *sses and what has RLHF to do with it

If you've ever asked ChatGPT, Claude, or Gemini a question and walked away thinking, "Wow, that was helpful," there's a good chance you were on the receiving end of something called RLHF. It's one of the quiet revolutions in modern AI: the

What is Google's KV Cache and will it resolve the Chip shortage?

What is Google's KV Cache and will it resolve the Chip shortage?

TurboQuant: Redefining AI efficiency with extreme compressionBoosting LLM Performance with Tiered KV Cache on Google Kubernetes Engine | Google Cloud BlogBoost LLM inference performance with LMCache on Google Kubernetes Engine. Discover how tiered KV cache expands NVIDIA GPU HBM with CPU RAM and local SSDs, significantly improving context length, concurrency, and

Kimi 2.6 is out , more powerful than ever

Kimi 2.6 is out , more powerful than ever

Kimi K2.6 Tech Blog: Advancing Open-Source CodingKimi K2.6 advances open-source coding, featuring long-horizon coding, coding-driven design, agent swarms, proactive agents, and the Claw Groups research preview.baseten.co logomoonshotai/Kimi-K2.6 · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open

Paperclip vs OpenClaw: Not Just Another Agentic Orchestration Tool

Paperclip vs OpenClaw: Not Just Another Agentic Orchestration Tool

Paperclip just crossed 31,000 stars on GitHub, and my timeline is full of people calling it "the future of AI companies." But scroll through the threads and you'll find a recurring question: isn't this just OpenClaw with extra steps? It's not.

A New Harness in Town: Meta-Harness

A New Harness in Town: Meta-Harness

Meta-Harness: End-to-End Optimization of Model HarnessesMeta-Harness automatically optimizes model harnesses — the code determining what to store, retrieve, and present to an LLM — surpassing hand-designed systems on text classification, math reasoning, and agentic coding.Yoonho LeeMeta-Harness: End-to-End Optimization of Model HarnessesThe performance of large language model (LLM) systems depends not only

What is Gemma 4 and how to use, finetune it

What is Gemma 4 and how to use, finetune it

Google just dropped Gemma 4, calling it their most capable family of open models to date. Built from the same research behind Gemini, these models pack serious multimodal intelligence into packages small enough to run on your phone and large enough to compete with frontier models on a server. If

How to finetune Yuan 3.0 on your local machine - Practical Guide

How to finetune Yuan 3.0 on your local machine - Practical Guide

We previously wrote about how to fine-tune Kimi 2.5 . We talked about Yuan 3.0 in depth in another post. This time we're tackling Yuan 3.0 Flash — a 40B-parameter MoE model that activates only 3.7B parameters per inference. It was built specifically for enterprise document

5 Hidden Easter Eggs in the Claude Mythos Preview System Card

5 Hidden Easter Eggs in the Claude Mythos Preview System Card

Anthropic just dropped the system card for Claude Mythos Preview — their most capable model to date, and one they've decided not to release publicly. At 245 pages, it's dense. Most commentary has focused on the big-picture safety story: the model is too capable for general release,

LlamaIndex LiteParse: Fast, Local Document Parsing for AI Agents

LlamaIndex LiteParse: Fast, Local Document Parsing for AI Agents

Every AI agent needs to read documents. Whether it's a coding assistant pulling context from a PDF spec, a research agent digesting a batch of papers, or an enterprise workflow processing invoices, document parsing is the unglamorous but essential plumbing that makes agentic AI actually work. The problem

What the heck is spec coding?

What the heck is spec coding?

While some of us are still trying to grasp what vibe coding means, a new concept just got dropped: spec coding. Well what the heck is spec coding and how does it compare to vibe coding? Is it just another loaded term for something simple? If it is useful, how

What is Karpathy's Autoresearch and how does it work?

What is Karpathy's Autoresearch and how does it work?

GitHub - karpathy/autoresearch: AI agents running research on single-GPU nanochat training automaticallyAI agents running research on single-GPU nanochat training automatically - karpathy/autoresearchGitHubkarpathy In early March 2026, Andrej Karpathy — former head of AI at Tesla, co-founder of OpenAI, and one of the most influential voices in deep learning — open-sourced

Self-Evolving Agents: Three Frameworks That Let Your AI Improve Itself

Self-Evolving Agents: Three Frameworks That Let Your AI Improve Itself

You ship an AI agent. It works great on the demo. Then real users show up, edge cases pile on, and suddenly you're spending your weekends manually tweaking prompts and patching logic. Sound familiar? Self-evolving agents are a new class of AI systems designed to fix this exact

Yuan 3.0, another Multimodal Model that can significantly change document AI

Yuan 3.0, another Multimodal Model that can significantly change document AI

Yuan 3.0 | 源3.0Yuan 3.0Yuan3.0 Flash: An Open Multimodal Large Language Model for Enterprise ApplicationsMascot Sammy2GitHub - Yuan-lab-LLM/Yuan3.0: Yuan3.0: Mixture-of-Experts (MoE) Language ModelYuan3.0: Mixture-of-Experts (MoE) Language Model. Contribute to Yuan-lab-LLM/Yuan3.0 development by creating an account on GitHub.GitHubYuan-lab-LLMYuanLabAI (YuanLabAI)Org profile

Deep Dive into OpenClaw's Agentic Orchestration: design patterns and philosophy

Deep Dive into OpenClaw's Agentic Orchestration: design patterns and philosophy

OpenClaw went from zero to 247,000 GitHub stars in about two months. Originally called Clawdbot, then Moltbot, and finally OpenClaw, Peter Steinberger's project became the poster child for autonomous AI agents in early 2026. But beneath the viral buzz and the lobster branding lies a genuinely interesting

From Prompt Engineering to Harness Engineering: The Three Eras of Building with LLM-based solutions

From Prompt Engineering to Harness Engineering: The Three Eras of Building with LLM-based solutions

If you have been building with large language models over the past couple of years, you have lived through a quiet revolution in how we think about this work. It started with prompt engineering, the art of asking the model the right question. Then it shifted to context engineering, the

From PPO to GRPO to DAPO: Understanding RL for LLMs and Every Training Parameter Explained

From PPO to GRPO to DAPO: Understanding RL for LLMs and Every Training Parameter Explained

In our last post, we walked through how to fine-tune your own LLM using GRPO, Common Crawl, and Unsloth. We covered the full pipeline from data to deployment, but we intentionally skipped over something important: what do all those training parameters actually mean? If you looked at the GRPOConfig block

We Open-Sourced Engram: A Brain-Inspired Context Database for AI Agents

We Open-Sourced Engram: A Brain-Inspired Context Database for AI Agents

Today we're open-sourcing Engram, a context database for AI agents. It's available now on GitHub under the MIT license. We built Engram because we kept running into the same problem across every AI agent project we worked on: context management is broken, and everyone is reinventing

Is LLM's Context Window is a Vanity Metric

Is LLM's Context Window is a Vanity Metric

We've been thinking about this for a while now. Every couple months some LLM provider puts out a press release — "Now supporting 1 million tokens!" Then 2 million. We're pretty sure we'll see 5 million token context windows by end of year.

How to tune your own LLM with GRPO, Common Crawl and Unsloth

How to tune your own LLM with GRPO, Common Crawl and Unsloth

If you’ve been following the AI space at all, you know that fine-tuning your own LLM is no longer a luxury reserved for teams with massive GPU clusters and seven-figure compute budgets. The tooling has matured to a point where you can fine-tune a capable model on your own