Hey, it’s Andreas.
Welcome back to Human in the Loop — your field guide to the latest in AI agents, emerging workflows, and how to stay ahead of what’s here today and what’s coming next.

This week:
→ Google’s Willow quantum chip achieves a 13,000× speed milestone
→ IBM open-sources CUGA — a generalist enterprise agent framework
→ Stanford’s iconic Deep Learning Course returns — led by Andrew Ng & Kian Katanforoosh

Plus: a deep dive into how browsers are quietly becoming the next critical layer in the AI stack — and what OpenAI’s new Atlas release signals for the future.

Let’s dive in.

Weekly Field Notes

🧰 Industry Updates
New drops: Tools, frameworks & infra for AI agents

🔹 Google’s Willow quantum chip achieves 13,000× speed milestone
→ Willow demonstrates quantum error correction at scale — 13,000× faster than classical compute for certain workloads. This could supercharge AI training and optimization in the years ahead.

🌀 LangChain & LangGraph hit v1.0 for production-ready AI agents
→ First two major versions of their open source frameworks just launched.

🌀 Anthropic releases Claude Code on the Web
→ Claude Code is now available in browsers, not just in IDEs, making it easier for more people to use.

🌀 IBM open-sources CUGA, a generalist enterprise agent framework
→ CUGA (Computer Using Generalist Agent) bridges research and enterprise reality. The project has just been open-sourced.

🌀 Manus 1.5 introduces one-prompt full-stack app builder
→ One prompt can now scaffold, design, and deploy a full-stack app.
Manus is extending it’s end-to-end software creation capabilities.

🌀 Brave warns of AI browser security flaw post-Atlas release
→ Shortly after OpenAI’s Atlas launch, Brave issued a warning about agent-based browser vulnerabilities — from cookie hijacking to permission overreach.

🌀 Meta lays off 600 AI staff for leaner, load-bearing teams
→ Meta trims its “ bloated AI division”, aiming for smaller, faster units that focus on efficiency and infra resilience.

🌀 OpenAI adds Company Knowledge to ChatGPT Enterprise
→ Enterprises can now securely inject proprietary data into ChatGPT for personalized reasoning — no retraining required.

🎓 Learning & Upskilling
Sharpen your edge - top free courses this week

📘 Stanford’s Deep Learning Course (Autumn 2025) returns — by Andrew Ng & Kian Katanforoosh
→ The legendary CS230 course is back and freely available on YouTube. Four updated lectures already out, covering the full deep learning lifecycle — from fundamentals to generative models. Still one of the best free ways to learn AI, now refreshed for 2025.

📘 IBM Tutorial — Set up an end-to-end use case for local AI agent with MCP server and watsonx Orchestrate
→ Hands-on guide on connecting a local MCP server to watsonx Orchestrate Developer Edition using Docker Compose.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 Stack AI on enterprise AI adoption
→ 99% of Fortune 500s now use AI — but most still chase impact, not experiments. Stack AI’s new paper maps 65+ real agent use cases proving where ROI actually happens.

🔹 Perplexity shares AI workflow playbook for professionals
→ A concise, tactical guide to building personal AI workflows for research, writing, and strategy. Worth bookmarking if you want to integrate AI into your daily routine.

🔹 Microsoft Copilot gets major Human-Centered AI update
→ Major Copilot Fall release focused on making AI more human-centered — with significant improvements in usability, and overall user experience.

🔹 Anthropic published a new study on personalities of AI models
→ New research stress-tested 300K value trade-off prompts across 12 frontier models — revealing distinct moral and behavioral patterns between Anthropic, OpenAI, Google DeepMind, and xAI.

♾️ Thought Loop
What I've been thinking, building, circling this week

Browsers are quietly becoming the most important layer in the AI stack — especially for AI agents.

This week, OpenAI launched ChatGPT Atlas — its own agentic web browser.
Atlas turns ChatGPT from a chat interface into a web-native agent that can browse, act, and execute directly inside pages. It’s a major step beyond conversation — toward autonomous, multi-step reasoning on the open web.

And it’s not a small shift.
The web browser remains the most universal, most used interface we have.
Most of our data, actions, and decisions still live inside it. Traditional chat interfaces can’t handle real workflows, while AI agents need context that’s persistent, real-time, and embedded.

Over 90 % of digital work still happens inside the browser — it’s where we read, write, buy, research, and collaborate.

But OpenAI isn’t alone. Four players are now shaping what the AI browser era might look like:

Even though Google Chrome still feels untouchable (and started similar capbilities), momentum is building currently. The next platform shift might not be a new app — it might be the browser itself.

Here’s why the browser could become the next AI operating system (especially for AI agents):

1. The browser is the last true interface layer.
→ It’s already where users work — making it the perfect environment for AI (agents) to act without app-switching friction.

2. Agents need context — and the browser has it.
→ Tabs, history, scroll behavior, and DOM data give AI (agents) everything they need to understand intent and act autonomously.

3. Memory lives best in the browser.
→ It spans your entire digital footprint, letting AI (agents) learn, retain, and assist across time and tools.

4. Browsing is the gateway to execution.
→ The browser isn’t just where people consume — it’s where they execute.
That makes it a powerful environment for AI agents to assist with task execution, directly where work happens.

🔧 Tool Spotlight
A tool I'm testing and watching closely this week

nanochat — NanoChat — Andrej Karpathy’s $100 end-to-end ChatGPT clone

Karpathy just open-sourced NanoChat: a minimalist, full-stack ChatGPT-style model you can train from scratch — tokenizer to UI — on a single 8×H100 node in ~4 hours. The entire pipeline runs for under $100, making it one of the most accessible ways to build your own ChatGPT from the ground up.

Highlights
→ Full pipeline: tokenize → pretrain → finetune → serve
→ One script (speedrun.sh) trains a 1.9B model in ~4h
→ Clean, hackable 8K-line codebase (PyTorch + Rust BPE)
→ Built-in web UI + automatic eval report
→ Designed for learning, tinkering, and LLM education

You can try it here: github.com/karpathy/nanochat