19 explained so far

AI and system design news, explained

When something changes in AI or system design, we run it ourselves and explain what it means for the systems you build, in plain words, with the sources linked.

Follow in your feed reader

17 Oct 2026 · 21 min read

NVIDIA Nemotron 3.5 Bolt and Super VL, explained: a 74B agent model in 22 GB, and what it needs to runOn 15 October NVIDIA released two open Nemotron 3.5 models for AI agents: Super VL, a 120B model for servers, and Bolt, a 74B model squeezed to 22 GB so it can run on one local GPU. I read the Bolt file's own header without downloading it, counted its 74 billion parameters and its 2-bit experts myself, and worked out what memory it really needs: more than the "24GB" in its name.Nemotron 3.5Nemotron 3.5 BoltNemotron 3.5 Super VLNVIDIAread the explainer 14 Oct 2026 · 16 min read

JEV-35B, explained: an open decision model you can download, what it needs, and a test of its prompt trickJEV-35B is a model on Hugging Face that never writes text. You give it a situation, one question and up to 256 answers, and it returns a probability for each answer in about a fifth of a second, by its maker's measurement. Unlike the hosted decision APIs from TypeSafe, OpenAI and Microsoft, the weights are free to download under Apache-2.0. It is too big for my 24 GB Mac, so I checked its own config files and tried its prompt layout on a small chat model: with that layout the small model changed 8 to 11 of 24 answers when I reversed the options.JEV-35BAutoTrustdecision modelsclassificationread the explainer 13 Oct 2026 · 16 min read

Intel GPU D-series (D270), explained: 480 GB of memory on one card, and what that buys you for AI inferenceIntel's new data center GPU, the D-series (formerly Crescent Island), puts up to 480 GB of memory on one air-cooled PCIe card, five times the 96 GB of the NVIDIA card Intel compares it with, at about the same memory bandwidth (how many bytes the chip can read each second): 1.5 TB/s. I measured the rule that decides what that means on my own Mac, then worked it out for the card: more users and longer conversations held in memory, not faster answers.Intel GPU D-seriesIntel D270Crescent IslandAI inference hardwareread the explainer 12 Oct 2026 · 14 min read

EmbeddingGemma 2, explained: one open model for text, code, pictures and sound, tested on a laptopOn 6 October 2026 Google released EmbeddingGemma 2, a free, open model that turns text, code, pictures, video and sound into lists of 768 numbers, so that things with similar meaning get similar numbers. That is the engine behind search by meaning and RAG. I ran it on a MacBook Air: it found the right text for 7 of 8 plain-English questions and the right picture for 6 of 6, at about 47 ms per sentence, and cutting its vectors to a sixth of their size kept the answers but squeezed the scores together.GoogleEmbeddingGemma 2embeddingssemantic searchread the explainer 12 Oct 2026 · 13 min read

Isometric drawing: the Claude Code and Codex coding agents send an MCP tool call to the REA server; REA drives Ghidra, Hopper or IDA for native binaries, a JavaScript parser for Electron and web code, and Chrome for live websites; the JavaScript parser reads the target app; a green dashed arrow carries evidence back to the agent

REA (Reverse Engineer Anything), explained: what the viral GitHub repo does, and what it missed in a real testREA is a free, open-source tool that lets an AI coding agent such as Claude Code or Codex look inside an app you do not have the source code for, and explain how a feature works. It went from 3 points on Hacker News in July to more than 109,000 GitHub stars this week. I ran it on a real open-source app: every place it pointed to was right, but it did not record any of the 8 places where the app sends a reply back to a window. Its own output warns that a missing finding proves nothing.REAreverse engineeringMCPAI agentsread the explainer 10 Oct 2026 · 12 min read

Microsoft-Decision-1, explained: a model that scores answers instead of writing them, and a test to runOn 9 October 2026 Microsoft released Decision-1, a model that never writes text. You give it some text, a question and a fixed list of answers, and it returns a probability for each answer. It costs $0.042 per million input tokens, output is free, and it runs in Microsoft Foundry and on OpenRouter. I could not call it myself, so I ran the test behind Microsoft's main robustness claim on two other models: qwen2.5:3b, a small chat model, changed its answer on 4 of 24 tickets when I reversed the options, and it was at least 99.9% sure both times.MicrosoftMicrosoft-Decision-1decision modelsclassificationread the explainer 9 Oct 2026 · 8 min read

OpenAI's Decisions API, explained: I ran it on 33 egg photos, and the top answer nearly binned a clean eggOpenAI's Decisions API, in public beta since 6 October 2026, does not write a reply. You send a photo or text, a question and a fixed list of answers, and it returns a probability for each. I ran it on 33 egg photos: it answered in about 0.2 seconds, and one clean egg came back 47% dirty, 44% clean. Here is what that means for your code.OpenAIDecisions APIclassificationconfidence thresholdsread the explainer 8 Oct 2026 · 19 min read

Claude Haiku 5.5, explained: a tenth of the price up to 100,000 tokens, and what breaks when you switchClaude Haiku 5.5 is Anthropic's new small, fast model, released on 7 October 2026. Up to 100,000 tokens of prompt it costs a tenth of Haiku 4.5 per token. I read Anthropic's pages, ran the same 12 support tickets through both models, and worked out where the price jumps and which old code stops working.Claude Haiku 5.5AnthropicClaudeLLM pricingread the explainer 7 Oct 2026 · 21 min read

DeepSeek V4.1 Flash, explained: an open model that does less work reading your prompt, and what it costsDeepSeek V4.1 Flash is a large AI model from the Chinese company DeepSeek that anyone can download under the MIT licence or call through DeepSeek's API for a low price. It came out on 10 September 2026. I read DeepSeek's own pages, ran its prompt code and tests on my Mac, and worked out the bill and the hardware.DeepSeek V4.1 FlashDeepSeekopen-weight modelsmixture of expertsread the explainer 7 Oct 2026 · 17 min read

Mistral Large 4, explained: a 1 trillion parameter model with open weights promised, and what it takes to runMistral Large 4 is a very large AI model that companies will soon be able to download and run on their own servers, if they can afford the machine. Mistral, a French AI company, opened an early preview on 6 October 2026 and promises the files by the end of October. I checked what I could reach today, ran the same code on a small Mistral model, and worked out the hardware and the bill.Mistral Large 4Mistral AIopen-weight modelsmixture of expertsread the explainer 6 Oct 2026 · 16 min read

The Codex agent loop in the middle: read the code, run a command, change a file, check the result. Around it, the places Codex runs: the CLI, the IDE extension, the ChatGPT desktop app, Codex Cloud and GitHub code review.

OpenAI Codex explained, and how it compares with Claude CodeCodex is OpenAI's AI coding agent: a program that reads your code, runs commands and makes changes on its own, in a terminal, a code editor, the ChatGPT app or OpenAI's cloud. This page explains what Codex is in October 2026, what it costs, and how it differs from Claude Code. I also tried it with a free local model and no OpenAI account. It found the bug, then failed to save the fix.CodexOpenAIClaude CodeAI coding agentsread the explainer 6 Oct 2026 · 14 min read

Three Qwen 3.8 open-weight models drawn as stacks of different heights: Qwen3.8-27B, 27 billion parameters, Apache 2.0, 18 GB in Ollama; Qwen3.8-Flash-Next, 125 billion with 6 billion active; Qwen3.8-2.4T-A95B, 2.4 trillion with 95 billion active. A label marks the 27B as the one we ran. Stack heights show order of size, not scale.

Qwen 3.8 explained: what is new, and running the 27B model with OllamaQwen 3.8 is the newest family of AI models from Alibaba's Qwen team. Three of its models have open weights: files you can download and run yourself. I ran the smallest one, Qwen3.8-27B, with Ollama, a free tool that runs models on your own machine, and recorded every step. This page covers what is new, what it needs, and what I measured.Qwen 3.8Qwenopen-weight modelsOllamaread the explainer 6 Oct 2026 · 19 min read

An AI app on the left with three plugs labelled client. Each plug connects to one MCP server on the right: a notes server, a GitHub server and a database server. Above the links: tools, resources and prompts. Below: one client per server, JSON-RPC over stdio or HTTP.

MCP explained: what the Model Context Protocol is, and a tiny server you can runMCP, the Model Context Protocol, is a free, public set of rules (a protocol) that lets an AI app use outside tools and data in one common way. A program that offers tools this way is called an MCP server. I built a tiny one on my Mac, connected to it, and recorded every message. This page explains MCP, what changed in the July 2026 version, and how to keep it safe.MCPModel Context ProtocolAI agentstool callingread the explainer 5 Oct 2026 · 19 min read

Three rings drawn on grid paper. In the centre, Claude Code. Around it, a ring labelled mods. Around that, a dashed ring labelled your machine. On the left, what a mod changes: draws panes and bands, rewrites a tool call, answers without the tool. On the right, what it reaches: your files and programs, the network and your keys, every prompt you send.

Claude Code mods explained: build your first mod, and what it can reachA Claude Code mod is a small piece of JavaScript or TypeScript that runs inside Claude Code, Anthropic's AI coding tool, and changes how it looks and what it does. It comes inside a plugin, an add-on package. I built one on my Mac, ran it, and read what a mod is allowed to touch.Claude CodeClaude Code modsAnthropicpluginsread the explainer 5 Oct 2026 · 16 min read

Three bars on a log scale of output price per million tokens: GPT-6 Luna $0.10 input and $0.50 output, GPT-6.1 Sol $2.00 and $10.00, GPT-6 Astra $10.00 and $50.00. Heading: GPT-6.1 Sol: 20x Luna's price, a fifth of Astra's. A note with my ticket test: median per call, measured 4 October, Sol 2.528 s and Luna 1.325 s; cost per 1,000 tickets, measured 5 October, Sol $1.0566 and Luna $0.0517.

ChatGPT 6.1 Sol explained: what GPT-6.1 Sol is, what it costs, and when to use itGPT-6.1 Sol is the middle of OpenAI's three GPT-6 models, released on 29 September 2026. You pay per token, a small piece of text: $2 for every million tokens you send and $10 for every million it writes. That is 20 times the price of Luna, the cheap model, and one fifth of Astra, the most capable one. I used it with my own OpenAI account and compared it with Luna on 300 support tickets.GPT-6.1 SolChatGPTOpenAILLM costread the explainer 4 Oct 2026 · 15 min read

A loop of four steps drawn on grid paper: read the files and run the tests, edit backoff.py, run the tests again, reply. In the middle: 4 turns, 8.9 seconds, an estimated $0.24. Heading: One prompt, four turns, tests green.

Claude Code, explained: what changed this week, and a real bug fix in four turnsClaude Code is Anthropic's AI coding agent. On 1 October 2026 it gained Claude Mods, a way to change how the tool itself behaves. I read the changelog and docs, then gave it a failing test on my own Mac and recorded every step, every token and the cost.Claude CodeAnthropicAI agentscoding agentsread the explainer 4 Oct 2026 · 12 min read

A request goes first to gemini-4-argon and comes back HTTP 404 in about 0.6 seconds, then to gemini-3.8-flash, which answers with HTTP 200 in about 2 seconds. Heading: Announced, but not on the API yet.

Gemini 4 Argon, explained: Google's new frontier model, and why I could not call it yetGoogle announced Gemini 4 Argon on 30 September 2026, with a 1 million token output limit and a price of $2 in and $10 out per million tokens. It is only open to a small group of cyber defenders for now, so I checked our own API key, got a 404, and turned that into a lesson on choosing models in production.Gemini 4 ArgonGoogleGemini APILLM pricingread the explainer 4 Oct 2026 · 11 min read

Two short lanes on grid paper under the heading: A chat waits. A dot keeps working. A chat: you ask, you leave, it waits. A dot: you give it a goal, you leave, it keeps working.

OpenAI dots, explained: how an always-on AI agent keeps working after you close the chatOn 29 September 2026 OpenAI launched dots: ChatGPT agents that have their own cloud computer and keep working after you leave. Here is what a dot is, who can use one, and how you would design an agent like this yourself.OpenAIAI agentsChatGPTdurable executionread the explainer 4 Oct 2026 · 12 min read

A large square stands for Kolibri's 78.1 billion parameters held in memory. A small square in its corner, drawn to scale by area, stands for the 3.46 billion parameters, 4.4 percent, used for each token. Heading: Kolibri computes like 3B but needs memory like 78B.

Kolibri, explained: Aleph Alpha's open model computes like a 3B model but needs memory like a 78B oneAleph Alpha released Kolibri on 3 October 2026: downloadable weights under Apache 2.0, built for German and English. It would not run on our 24 GB Mac, which turned out to be the most useful lesson, so we ran its tokenizer instead and worked out what it needs from your hardware.KolibriAleph Alphaopen-weight modelsmixture of expertsread the explainer