Subscribe, if this was forwarded to you

Hey there,

This week, Google's Gemini 4 Argon is open to paying developers, built for long coding and agent runs with up to 1M output tokens. The tutorial is Leiden's free graduate textbook on how agentic LLMs are built, and DigUp searches the photos, PDFs and recordings on your own computer.

We'll also look at Muse Spark Open and Mistral Large 4, OpenAI's answer to Jev, why Anthropic replaced git, Claude Haiku 5.5's new price, and a tip for making agent output shorter.

Enjoy the read!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: A Free Graduate Textbook on Agentic LLMs

🛠️ Tool Spotlight: DigUp: Search Your Files by What's in Them

📰 Top News: Gemini 4 Argon Opens to Developers

💡 Tip: Make Agent Output Shorter with Controlled English

Let's get started!

Tutorial of the week

A Free Graduate Textbook on Agentic LLMs

Agent frameworks ship faster than explanations of how an agent is actually put together. Leiden University's free textbook, Agentic Large Language Models, is the text for a master's course of the same name. The authors are Max van Duijn, Michiel van der Meer, Aske Plaat and Niki van Stein.

• Start from a definition: an agentic LLM is a model plus a harness, the code that builds prompts, calls tools, keeps state, and loops through reasoning, acting and observing.

• Cover the whole stack: seven parts run from language modeling and reasoning through agents, tools and assistants, behavior and cognition, and advanced topics.

• Use it as a course: Leiden teaches it as 14 lectures and four graded case studies, with no separate theory exam. Slides and assignments are at agenticllms.liacs.nl.

• Check the prerequisites: undergraduate Python, PyTorch, machine learning, some reinforcement learning and transformers. The book reviews the essentials if a piece is missing.

If you want a structured course on how agentic systems are built, rather than another framework tutorial, this is a free place to start.

Tool of the week

DigUp: Search Your Files by What's in Them

Search on your computer finds a file by its name. DigUp, an MIT-licensed desktop app, finds it by what is in it: a photo, a page in a PDF, a moment in a recording, or a frame in a video. It runs Google DeepMind's EmbeddingGemma 2, a 740-million-parameter Apache 2.0 model that puts text, images, audio and video in one space and covers more than 100 languages.

• Describe it: "zebra in a video" opens the clip at that moment, and "where they talk about sleep" jumps to that point in a recording.

• Keep it on the machine: nothing you index is uploaded. DigUp goes online only to download the model once (865 MB) and to check for updates, which you can turn off.

• Skip what you should not index: keys, certificates, hidden folders such as .ssh, and files that exist only in the cloud. Code search stays off until you turn it on, and then the query starts with code:.

• Open the hit: ⇧⌘Space on a Mac, Ctrl+Shift+Space on Windows and Linux. A PDF opens at the page, a video at the moment, code at the line. Mac needs Apple Silicon and macOS 14 or later. Windows and Linux are still in beta.

If your folders are full of screenshots, PDFs and recordings that search-by-name never quite finds, this is a local way to look.

Top News of the week

Gemini 4 Argon Opens to Developers, with Up to 1M Output Tokens

On 13 October, Google opened Gemini 4 Argon to paying developers, three weeks after announcing it only for trusted cyber defenders in the Fairwind Program. The model id is gemini-4-argon, on the paid Gemini API, in Google AI Studio and on Google Cloud. It is not on the free tier. Input, thinking and output share one window of 1,048,576 tokens, and a response can run past the old 64K output cap.

Google's introductory price runs through 30 November: $2 per million input tokens and $10 per million output tokens, thinking included. On 1 December that becomes $4 and $20. Cached input is 95% off until then, $0.10 per million.

Google reports 77.9% on DeepSWE v1.1 and first place on Zapier's AutomationBench at 51.3%. Inside Google, Argon agents freed over 300 TiB of data-center memory, and on libgav1 they produced Rust that runs 2.7× faster than Google's existing Rust port, with the same video output. A larger migration, more than 800,000 lines of the Fuchsia Zircon kernel, is still under audit. On the public API, Google says it can stop a run by monitoring the chain of thought. Fairwind defenders get a build without cyber guardrails.

The introductory price ends on 30 November, and the internal results are Google's own.

Muse Spark Open's 816B Weights Are Out; Mistral Large 4's Are Not

Meta released Muse Spark Open on 15 October: 816 billion parameters, 42 billion active, and a 1 million token context. NVFP4 needs about 700 GB of VRAM, or two M5 Ultra Mac Studios. Products must display "Built with Muse Spark Open," and an API business above $20 million in revenue needs a deal with Meta. Mistral Large 4 (1 trillion parameters, 52 billion active) is an API preview only, with weights promised by the end of October.

🔗 Muse Spark Open on Hugging Face

🔗 Mistral Large 4 announcement

OpenAI Takes On Jev with Its Decisions API

Two weeks after TypeSafe's Jev, our last top story, OpenAI put its Decisions API into public beta. It asks the same three kinds of questions as Jev: a yes/no probability, one choice from a list, or a score. OpenAI says it is about 10× faster than its Responses API. Unlike Jev, it also reads images and can refuse a question. It runs only on gpt-6-luna and costs $0.10 per million input tokens, against Jev's $0.042, and neither charges for output.

🔗 Decisions API guide

OpenAI Takes On Jev with Its Decisions API

Two weeks after TypeSafe's Jev, our last top story, OpenAI put its Decisions API into public beta. It asks the same three kinds of questions as Jev: a yes/no probability, one choice from a list, or a score. OpenAI says it is about 10× faster than its Responses API. Unlike Jev, it also reads images and can refuse a question. It runs only on gpt-6-luna and costs $0.10 per million input tokens, against Jev's $0.042, and neither charges for output.

🔗 Decisions API guide

Claude Haiku 5.5 Costs About 90% Less on Short Prompts

Anthropic priced Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 for output on prompts up to 100,000 tokens, against $1 and $5 for Haiku 4.5. That band was about 90% of previous Haiku requests, so Anthropic puts the average saving at about 75%. Above 100,000 tokens the cut is 50%. The same day, Sonnet 5.5 cache reads fell from $0.20 to $0.10 per million tokens.

🔗 Haiku 5.5 announcement

The Lean File Does Not Match OpenAI's Navier–Stokes Write-Up

A preprint from King's College London and Cambridge says the Lean formalisation of OpenAI's announced Navier–Stokes proof does not match the written argument. Lean checked a different claim. The authors do not say the result is false. A compiling Lean file is not, by itself, evidence that the prose proof is right.

🔗 Read the arXiv paper

Tip of the week

Make Agent Output Shorter with Simplified Technical English

Agent status lines and procedures often take a paragraph to say one action. Before: "Prior to commencing the installation, it should be ensured that all components have been thoroughly inspected for damage." After: "Before you start the installation, examine all the components for damage."

Simplified Technical English is the controlled language in the aerospace specification ASD-STE100: a fixed word list and short sentences, so a procedure stays unambiguous. The repository 0xpili/simplified-technical-english packages 53 writing rules, 869 approved words and a checker as an agent skill. It is a community project, not an official ASD document, and the dictionary itself belongs to ASD.

The practical lesson is to constrain what the agent writes back, not only what you ask:

• Install it or paste it: clone the repo into ~/.claude/skills/simplified-technical-english, or copy SKILL.md into another model's system prompt.

• Rewrite the noisy lines: ask the agent to put status updates and procedure steps into STE.

• Check before you ship: python3 scripts/ste_check.py --mode procedural draft.txt flags long sentences, the passive voice and words off the list.

• Keep a person on important text: the checker cannot tell whether a word is used in its approved sense. Use STE for procedures and status, not for chat or marketing copy.

If your agents write instructions that people have to follow, this is a small constraint with a visible before and after.

Read the Simplified Technical English skill

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help