FOLLOW THE
TERN.
CHASE THE FUTURE.

The world's best AI papers, books, podcasts, and tutorials — carefully curated from every perspective. Plus practical guides to building private LLMs, AI agents, and the next generation of intelligent systems.

The AI Hub Build Your Own AI Create AI Agents

The Week's Top 5 AI Stories — Most Read

Week of October 6, 2026 · ranked by reads

1Google unveils Gemini 4 Argon, retaking the benchmark lead over OpenAI and Anthropic — but in limited releaseVentureBeat
2FTC is investigating OpenAI, Anthropic and other AI companies over product risksCNBC
3OpenAI will start watermarking ChatGPT’s text in the EUTechCrunch
4Anthropic expands Claude Startups program in bid to snag founders and fast-growing companiesCNBC
5Can Google’s new model really catch up to OpenAI and Anthropic at the frontier?CNBC

The Week's Top 5 Health AI Stories

Week of October 6, 2026 · medicine & healthcare

Data Centers — Truths & Myths

Week of October 6, 2026 · environmental impact, pro & con

AI Jobs — Losses & Gains

Week of October 6, 2026 · the changing labor market

FRONTIER POWER RANKINGS

The Frontier Seven

The seven most powerful AI models on Earth — closed and open — scored on Edge + Alpha: raw intelligence, real-world usage, and profitability.

Closed source Open source Score = Edge + Alpha (0–100)

1GPT-6 AstraOpenAI · USA · Closed

Released September 3: OpenAI's most capable model yet and a new state of the art for computer and browser use. On OpenAI's own numbers it posts 97.6% on FrontierMath Tier 4, 74.1% on DeepSWE and 100% on ExploitBench. Controversial, too — its "opaque recurrence" reasoning makes chain-of-thought monitoring harder.

2Claude Opus 5.5Anthropic · USA · Closed

Released September 22 and called by Anthropic “the strongest-performing model we’ve tested to date,” outpacing the larger Fable model on many benchmarks. Output tokens drop to $20 per million from $25, and the model runs faster on less compute. The first Anthropic release since CEO Dario Amodei embraced pacing the frontier — alignment progress first, capability second.

3Gemini 4 ArgonGoogle · USA · Closed

Announced September 30 and, by Google’s own numbers, leading or tying on 13 of 18 disclosed benchmarks against GPT-6 Astra and Claude Opus 5.5 — though Astra still leads on software engineering. Rollout is limited to trusted cyber defenders for now, with introductory pricing that undercuts Astra by 80%.

4Kimi K3Moonshot AI · China · Open

Now fully open: Moonshot released the weights of the largest open model ever — 2.8T parameters, 1M-token context, Opus-class coding and agents. A landmark moment for open AI.

5DeepSeek V4.1 FlashDeepSeek · China · Open

The cost disruptor, now the volume leader: V4.1 Flash shipped earlier this month and sits at number one on OpenRouter’s leaderboard after a 172% jump in usage in a single week. Real-world adoption, not just a benchmark line — pure alpha.

6Qwen 3.7 MaxAlibaba · China · Open

The most-run open model on Earth — permissive license, every size, the default local brain for builders worldwide.

7Grok 4.6xAI · USA · Closed

Shipped August 12, just 35 days after 4.5: a 500K-token context tuned for long-running agents, now matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index at roughly half the price. Fast, cheap and wired to live data.

Featured Clip

The Day After AGI

Demis Hassabis (Google DeepMind) & Dario Amodei (Anthropic) debate what comes after AGI, moderated by The Economist’s Zanny Minton Beddoes · World Economic Forum

Every model in the global top seven is from the U.S. or China. The strongest outside them is Mistral (France — Europe's frontier lab), followed by Falcon (UAE), Cohere (Canada) and AI21 (Israel) — all just short of this list.

Composite of intelligence, adoption & profitability · September 2026 · for orientation, not investment advice.

Exciting news in the works at TERNZ AI…COMING SOON!

AI Hub

All you need to know about AI

Build Your AI

A step-by-step path to your own private model

Gear Up

The one number that decides everything is memory — it caps how big a model you can hold. Buy the most VRAM (Nvidia) or unified memory (Apple / AMD) your budget allows. A used 24GB card is the cheapest serious entry; a 48GB Mac is the comfortable sweet spot; 128GB runs frontier-class models at home. Tap a card to buy it.

RTX 3090 · 24GB · ~$700 RTX 5060 Ti · 16GB · ~$450 Mac Mini M4 Pro · 48GB · ~$1,800 DGX Spark · 128GB · $4,699

Hardware guide →

Install a Runtime

This is the software that loads a model and runs it on your hardware. Ollama is the fastest start — one command, works like Docker for models. LM Studio gives you a friendly desktop app with no terminal at all. llama.cpp is the bare-metal engine underneath both, for maximum control.

Ollama LM Studio llama.cpp

Setup docs →

Pull a Model

Open-weight models are ones you download and own outright — no API key, no logging, no cloud. Match the model to your memory: 7–14B for a laptop, 32–70B once you have 24–48GB. Check the live leaderboards to grab the current best for your size.

Qwen3 DeepSeek-R1 Llama 4 gpt-oss

Compare models →

Add an Interface

A runtime on its own is just a command line. Open WebUI puts a clean, familiar chat window on top — chat history, file uploads, multiple users — so your whole household gets a private ChatGPT that never touches the cloud.

Open WebUI

Interface docs →

Feed It Your Data

Retrieval-augmented generation (RAG) lets the model answer from your own documents without retraining it. Point a tool at your files; it indexes them into a local vector database and feeds the right snippets to the model at query time — all on your disk.

AnythingLLM LlamaIndex Chroma

RAG docs →

Give It Tools

An agent is a model that can take actions — search the web, run code, call APIs — not just chat. The Model Context Protocol (MCP) is the emerging open standard for plugging tools into any model; frameworks like LangChain wire the loop together.

Model Context Protocol LangChain

MCP docs →

Go Sovereign

Sovereignty means your data and your thinking never leave your control. Keep the sensitive machine offline or firewalled, encrypt the disk, switch off every telemetry setting, and keep your own copies of the model weights. Now nothing you do is logged, sold, or dependent on anyone else's servers.

Air-gap itEncrypt the diskKill telemetryBack up your weights

Privacy guide →

Scale It

When one chat window isn't enough — multiple users, an app, batch jobs — you need a real inference server. vLLM serves open models with high throughput and an OpenAI-compatible API, so your local rig can power a whole team or product.

vLLM

Serving docs →

Your Own Private AI

Yours. Local. Answering to no one but you.

Create AI Agents

An agent is AI that doesn't just answer, it acts: give it a goal and it plans, grabs tools, checks its own work, and keeps going until the job's done. Less chatbot, more tireless intern. Let me show you how they work — and how to build your own.

What is an Agent?

An agent has a goal, not just a prompt. It plans the steps, uses tools — web, code, files, APIs — observes what happened, and loops until it's finished. You hand over the outcome; it works out the how.

What is a Sub-Agent?

Big jobs overwhelm one brain. A lead agent splits the goal and hands each piece to a specialized sub-agent — one researches, one writes, one checks — then merges the results. Each stays focused; the lead keeps the master plan.

How to Build an Agent

Seven moves from idea to a working agent.

Set the goal & guardrails — what "done" is, and what it must never do.
Pick the brain — a model to reason and decide (see below).
Give it tools — web, code, files, APIs.
Add memory — so it recalls context and past steps.
Let it loop — plan → act → observe → repeat.
Split into sub-agents — hand big jobs to focused specialists.
Test & watch — trace it, set limits, iterate.

Closed-Source Brains

The most capable models today — best raw performance.

Claude — AnthropicTop tool-use & coding; a favorite for agents GPT — OpenAIBest all-rounder, biggest ecosystem Gemini — GoogleStrong reasoning, great value Grok — xAIFast, with live data

Open-Source Brains

Own it — run it on your own hardware, private and free.

Qwen3 — AlibabaApache 2.0, every size, top open pick DeepSeekBest open reasoning & math Llama — MetaHuge ecosystem & tooling gpt-oss — OpenAIOpenAI-quality, open weights MistralLean, permissive, reliable

Frameworks

Don't wire it from scratch — these handle the loop, tools and sub-agents.

LangGraphStateful graphs, human-in-the-loop CrewAIRole-based multi-agent crews, fast OpenAI Agents SDKLightweight single/few-agent path Claude Agent SDKNative sub-agent spawning Google ADKHierarchical agent trees AutoGenConversational multi-agent LlamaIndexRAG-grounded agents MCPOpen standard for tools

Learn It — Tutorials

Go deeper with hands-on guides from the people who build agents.

Building Effective AgentsAnthropic — the essential primer AI Agents CourseHugging Face — free, hands-on LangChain AcademyBuild & ship real agents Agentic Short CoursesDeepLearning.AI

The week's top AI, by email.