A model learns to use a computer from 18 years of video, Sakana tops real cyber-defense benchmarks, a frontier world model gets a full open reproduction — and World models and security, both getting real. Datalab's Marker 2 vs MinerU, Docling & LiteParse: The Doc-Parsing Benchmark
Sakana AI launched Fugu-Cyber, a multi-agent orchestration model for cybersecurity that packages a whole agent system as one model. It hits 86.9% on CyberGym and 72.1% on CTI-REALM — topping GPT-5.5-Cyber and Mythos-Preview on both.
So what: "Multi-agent system as a model" is the interesting move — orchestration baked in, not bolted on. Beating dedicated cyber models from OpenAI-class labs on real benchmarks signals security is now a distinct model category where specialist orchestration wins.
A head-to-head of open PDF-to-markdown parsers. Marker 2 scores 76.0 on olmOCR-bench at 2.9 pages/sec — matching Gemini/MinerU quality while running ~5x faster, beating MinerU and Docling on score and throughput at once.
So what: Document parsing is the silent failure point of most RAG stacks — bad extraction poisons everything downstream. Winning quality and speed simultaneously makes Marker 2 the new default front door for turning PDFs into clean model input.
sponsored
Induction Labs' first "imagination model," Photon-1, learned to use a computer by watching the equivalent of 18 years of screen recordings — with no action labels. From a single pretraining run it can simulate desktops, play tournament checkers (after fine-tuning on 20,000 games), and model billiard physics.
So what: One unlabeled-video pretraining run yielding desktop control, game play, and physics is the world-model thesis paying off — learn the dynamics once, specialize cheaply. If screen-recording video is enough to learn computer use, the data moat for agents just got a lot shallower.
Open-Dreamer reproduces the Dreamer 4 world-model pipeline in JAX/Flax NNX and — crucially — publishes the full training recipe: code, configs, and the stability-engineering lessons. Trained on Minecraft/VPT-style data, it's a clean, performant open implementation.
So what: World models have been reproducibility-hostile — the training tricks live in a few labs' heads. Publishing the recipe, not just weights, is how the technique actually spreads. This is the reference that turns Dreamer 4 from a paper into something you can train.
A hands-on build: a stateful agent that runs a venue-operations workflow end to end. MongoDB Atlas for data + vector search, Voyage embeddings for retrieval, and LangGraph to orchestrate the planning / tool-calling loop into a durable graph.
So what: A clean reference for the emerging "database + embeddings + graph orchestration" agent stack — the pattern most production agents are quietly converging on.
sponsored
FEATURED TUTORIALS with CODES
|
Embeddings EmbeddingsHow AI Turns Meaning Into Math |
|
Panel 1 · The Problem
|
|
Panel 2 · The Big Idea
Modern AI uses 768 to 1,536 numbers per word. No single number means "animal" or "size" — but together, they capture meaning. How? By reading billions of sentences. "Dog" and "puppy" both appear near "bark" and "leash" — so they must be similar! (Word2Vec, 2013 💵) |
|
Panel 3 · 🕷 The Aha Moment King − Man + Woman = QUEEN 👑 Nobody taught the AI what "gender" is. |
|
|
Panel 4 · Why You Should Care
|
|
The Takeaway Embeddings turned meaning into math. |
$ ./partner --with us
# Reach 1M+ AI developers, researchers & builders.
# Sponsor a future issue of the Intelligence Brief.
> apply: forms.gle/CY1eqZzuWFQBp7dH9